A closer look at referring expressions for video object segmentation

The task of Language-guided Video Object Segmentation (LVOS) aims at generating binary masks for an object referred by a linguistic expression. When this expression unambiguously describes an object in the scene, it is named referring expression (RE). Our work argues that existing benchmarks used for LVOS are mainly composed of trivial cases, in which referents can be identified with simple phrases. Our analysis relies on a new categorization of the referring expressions in the DAVIS-2017 and Actor-Action datasets into trivial and non-trivial REs, where the non-trivial REs are further annotated with seven RE semantic categories. We leverage these data to analyze the performance of RefVOS, a novel neural network that obtains competitive results for the task of language-guided image segmentation and state of the art results for LVOS. Our study indicates that the major challenges for the task are related to understanding motion and static actions.


Introduction
Video Object Segmentation (VOS) [33,44] has been traditionally considered on setups where a user would annotate the pixels of an object in a video frame, and an automatic system would extend this to the rest of video frames where the object is visible. Our work aims at simplifying the human-computer interaction by allowing linguistic expressions as initialization cues, instead of user interactive segmentations under the form of a detailed binary mask, bounding box, scribble or points. In particular, we focus on referring expressions (REs), which allow the identification of an individual object in a discourse or scene (the referent). REs unambiguously identify the target instance. For example, Fig. 1 depicts REs related to one of the objects contained in a video sequence, which is highlighted in green. Video sequences for DAVIS-2017 with REs and our results. The first column shows a reference frame, the second to third columns depict the masks produced by our model when given the RE shown on top. Finally, the fourth to fifth columns show the results for the REs shown on top of these columns, which refers to another object of the video sequence Language-guided Video Object Segmentation (LVOS) was first addressed by Khoreva et al. [21], and later tackled by Gavrilyuk et al. [12] and Wang et al. [40]. Compared to related works on still images [4,48], REs for video objects may be more complex, as they can refer to variations in the properties of the objects, such as a change of location or appearance. The particularities of REs for videos were initially addressed by Khoreva et al. [21], who built a dataset of REs divided in two categories: REs for the first frame of a video, and REs for the full clip. Our work proposes another approach for analyzing the performance of the state of the art in LVOS. We identify seven categories of REs, shown in Table 1, and use them to annotate existing datasets.
The main goal of our work is to study the task of LVOS, by focusing on the effect the different categories of REs have on the model performance in current benchmarks. We address both the language-guided image segmentation and the LVOS tasks with RefVOS, our end-to-end deep neural network that leverages the language representation model BERT [10] to encode the phrases into distributed representations. RefVOS stands for Referring Expressions for Video Object Segmentation. Our model achieves results comparable to the state of the art for the RefCOCO dataset of still images [20], and improves the state of the art over the DAVIS-2017 [34] and Actor-Action datasets (A2D) [43] for video datasets augmented with the phrases collected by Khoreva et al. [21] and Gavrilyuk et al. [12], respectively. We also identify the categories of REs which are most challenging for RefVOS, using our own REs annotations for the A2D dataset.
Our main contributions are summarized as follows: (1) an end-to-end model, namely RefVOS, that achieves state of the art performance with available expressions for DAVIS-2017 and A2D benchmarks, (2) a novel categorization of REs tailored to the video scenario with an analysis of the current benchmarks, and (3) an extension of A2D with additional REs of varying semantic information to analyze the limitations and strengths of our model according to the proposed linguistic categories.
The models, code and extended dataset of REs are available at https://github.com/ miriambellver/refvos.

Language-guided image segmentation
The task of Language-guided Image Segmentation, also known as referring image segmentation, was first tackled by Hu et al. [16]. They used VGG-16 [39] to obtain a visual representation of the image, and a Long Short-Term Memory (LSTM) network to obtain an embedding of the RE. From the concatenation of visual and language features, the segmentation of the referred object is obtained. Posterior work [26] explored how to include multi-scale semantics in the pipeline, by proposing a Recurrent Refinement Network that takes pyramidal features and refines the segmentation masks progressively. Liu et al. [27] argued to better represent the multi-modality of the task by jointly modeling the language and the image with a multi-modal LSTM that encodes the sequential interactions between words, visual features and the spatial information. With the same purpose of better capturing the multi-modal nature of this task, long-range correlations between the visual and language representations were reinforced by learning a cross-modal attention module (CMSA) [46]. Building on the same idea, BRINet [17] added a gated bidirectional fusion module to better integrate multi-level features. STEP [4]. learned a visual-textual co-embedding that iteratively refines the textual embedding of the RE with a Convolutional Recurrent Neural Network, in a collaborative learning setup to improve the segmentation. An alternative may consist of using off-the-shelf object detectors, like MAttNet [48]. In this case, a language attention network decomposed REs into three components: subject, location, and relationships, and merged the features obtained for each into single phrase embeddings. Given the object candidate by the off-the-shelf object detector model and a RE, the visual module dynamically weighted scores from all three modules to fuse them. A different approach was proposed in CMPC [18], which leveraged multi-modal graph reasoning to identify the target objects.
Whereas previous works mainly focus on how to better exploit the REs by designing language encoders tailored to the task, our proposed architecture RefVOS leverages BERT [10] to obtain the language representations. BERT is a bidirectional language encoder based on transformers, a type of neural network architecture that is used by current state of the art models in natural language processing [35] and are also more and more used in computer vision [2,42]. BERT is thus a strong baseline for obtaining language embeddings. Another characteristic of our work, compared to previous methods such as MAttNet [48], is that RefVOS directly produces pixel-wise annotations, without requiring object detections, which allows to train our model end-to-end. As a visual encoder, we use DeepLabv3 [5], a state-of-the-art model for segmentation. Compared to other works which perform early-fusion of multi-modal features (CMPC [18]), we apply late-fusion of the visual and language features in order to predict the final segmentation of the referred region. Thanks to late-fusion, we can work with off-the-shelf models for both the visual and the language encoders. As a conclusion, RefVOS is a simpler model trained end-to-end that obtains a performance comparable to the state of the art on still images.

Language-guided video object tracking
Object Tracking is a task similar to Video Object Segmentation as it also follows a referent across video frames, but in the tracking case the model localizes the object with a bounding box instead of a binary mask. Li et al. [24] and Feng et al. [11] tackle the object tracking problem given a linguistic expression instead of providing a bounding box for the first frame.
Our work provides pixel-wise segmentation masks that could be easily converted into bounding boxes, and at the same time avoid the annotation ambiguities that result from overlapping bounding boxes.

Language-guided video object segmentation (LVOS)
Video Object Segmentation (VOS) [33,44] has traditionally focused on semi-supervised setups in which a binary mask of the object is provided for the first frame of the video. Khoreva et al. [21] proposed to replace the mask supervision with a linguistic expression. In their work, they extended the DAVIS-2017 dataset [34], a popular dataset for VOS, by collecting REs for the annotated objects. They provide two different kinds of annotations collected by two annotators: first frame annotations are the ones that are produced by only looking at the first frame of the video, whereas full video annotations are produced after seeing the whole video sequence. They used the image-based MAttNet [48] model pre-trained on RefCOCO to ground the localization of the referred object, and then trained a segmentation network with DAVIS-2017 to produce the pixel-wise prediction. To ensure coherent bounding boxes across frames, they enforced temporal consistency with a post-processing step. To the authors' knowledge, Khoreva et al. [21] is the only work prior to ours that focuses on REs for VOS. Related work by Gavrilyuk et al. [12] addresses a similar task by segmenting video objects given an expression. They extend the Actor-Action Dataset (A2D) [43] by collecting linguistic expressions, but some of them may be ambiguous with respect to the intended referent, as they were not produced with the aim of unique reference, but description. The authors propose a model with a 3D convolutional encoder and dynamic filters that specialize in localizing the target objects. Wang et al. [40] also leveraged 3D convolutional networks, adding cross-attention between the visual and the language encoder. Concurrent to our work, Seo et al. [38] proposed URVOS, a model for LVOS composed of a cross-modal attention module for the visual and language features, and a memory attention module to leverage information from past predictions in a sequence.
Compared to the aforementioned methods, our work proposes a simpler model trained end-to-end that treats each video frame independently. Hence, RefVOS does not exploit the temporal information, yet it outperforms all previous works. Our method has the advantage that it can be used on either still images or video sequences, and it does not rely on heavy 3D encoders such as the architectures by Gavrilyuk et al. [12] and Wang et al. [40]. In contrast to Khoreva et al.'s [21] approach, our network can be trained end-to-end as it does not rely on any detection network or post-processing step. Finally, the most recent work URVOS [38] is a complex architecture due to the cross-attention and memory network. RefVOS proposes a much simpler architecture composed of two independent encoders with late-fusion of embeddings.

Categorization of referring expressions
RefCOCO, RefCOCO+ [49] and RefCOCOg [30] are datasets that provide REs for the still images in the MSCOCO dataset [25]. Each dataset focuses on different aspects related to the difficulty of REs: those of RefCOCO and RefCOCO+ were collected using the interactive ReferIt two-player game [20], designed to crowdsource expressions that uniquely identify the target referents. However, for RefCOCO+, location information was disallowed. Ref-COCOg, in turn, collected non-interactively, only contains non-trivial instances of target objects, that is, there is at least one other object of the same class in an image (e.g., multiple dogs). The CLEVR [19] and CLEVRER [47] datasets contain objects of certain shapes, attributes such as sizes and colors, and spatial relationships. CLEVR uses synthetic images and linguistic expressions designed to test visual question answering systems, while our work focuses on human-produced language and natural videos. CLEVR-Ref+ [29] extends the CLEVR dataset by adding REs for Language-guided Image Segmentation instead of visual question answering, but using the same synthetic images.
Khoreva et al. [21] categorize the REs they collected for DAVIS-2017 in order to analyze the effectiveness of their proposed model. This is similar to our work, however, while they distinguish REs according to their length and whether they contain spatial words (e.g., left) or verbs, we propose a more fine-grained, semantic categorization, presented in Table 1, that also distinguishes between different aspects of verb meaning related to motion and object relations. Khoreva et al. [21] further analyze the REs in DAVIS-2017 with respect to the parts of speech they contain, while we use our semantic categories for dataset analysis.

RefVOS model
The task of language-guided image segmentation is to, given a still image and a linguistic expression, segment the region to which the expression refers. Language-guided video object segmentation (LVOS) is a natural extension of this task, which aims at segmenting the referred object in the different video frames of the sequence. The latter is a more complicated task as the linguistic expression can refer to the motion or scene changes in the video.
In our work, we address the task of language-guided image segmentation and LVOS with the deep neural network depicted in Fig. 2, that we call RefVOS. RefVOS operates at the frame level, i.e., it treats each frame independently, and is thus applicable to both images and videos. It uses state of the art visual and linguistic feature extractors, which are combined into a multi-modal embedding decoder to generate a binary mask for the referent.
In order to define the architecture of RefVOS, we first formalize the LVOS task: In this task, the input of the system (that we define as X) consists of an input video (V ) and a linguistic expression (L), i.e., X = {V , L}. Each video sequence is composed of T video frames V = {I 0 , I 1 , I 2 , . . . I T −1 }, and each linguistic expression is composed of M tokens L = {w 0 , w 1 , w 2 , . . . w M−1 }. Note that only video frame I t is different at each time step t, whereas the expression L is kept constant for all video frames.  Fig. 2 Architecture of our RefVOS model

Visual encoder
To encode the images, we rely on DeepLabv3, a network for semantic segmentation based on dilate / atrous convolutions [5]. We use DeepLabv3 with a ResNet101 [14] backbone and an output stride of 8. The Atrous Spatial Pyramid Pooling (ASPP) has atrous convolutions with rates of 12, 24 and 36. Following the formulation defined in Section 3, the visual embedding is obtained independently for each video frame (I t ). In the case of the LVOS task, for each video frame I t ∈ V , we define the visual embedding as e v,t = V enc (I t ), where V enc () is the DeepLabv3 network.

Language encoder
In contrast to previous works addressing language-guided image segmentation, we are the first ones to leverage the bidirectional transformer model BERT [10] as language encoder. We use BERT in our pipeline in order to obtain an embedding for the linguistic expressions. We choose the BERT-base architecture, which is composed of 12 layers and a total of 110M parameters. We first fine-tune BERT (originally pre-trained on the BookCorpus [52] and the English Wikipedia) with the REs of RefCOCO using the masked language modelling (MLM) loss for one epoch, which consists of randomly masking a percentage of input tokens and then predicting them, following the common fine-tuning procedure for BERT. We then integrate BERT into our pipeline and fine-tune it specifically towards the language-guided image segmentation task: to this end we tokenize the linguistic expression L and add [CLS] and [SEP] tokens at its beginning and end, respectively. BERT produces a 768-dimensional embedding for each input token. We adopt the procedure of Devlin et al. [10] and extract the embedding corresponding to the [CLS] input token, i.e., the pooled output, as it aggregates a representation of the whole sequence. The encoded expression is then converted to a 256-dimensional embedding with a linear projection. We formally define the output of the language encoder as e l = L enc (L), where L enc () first tokenizes the linguistic expression L into the sequence of tokens {w 0 , w 1 , w 2 , . . . w M−1 }, then obtains the pooled output from BERT-base, and finally produces the 256-dimensional language embedding e l . Note that this embedding is unique for the whole video sequence V .

Multi-modal embedding
At each time step t, we obtain a multi-modal embedding Y t by performing element-wise multiplication of the language features and the visual features extracted by the ASPP from DeepLabv3. That is, Y t = e v,t ⊗ e l . 1 A convolutional layer then predicts two maps, one for the foreground and another one for the background class, S t = F (Y t ), where S t is the segmentation result at time step t and F is the convolutional layer. We employ the crossentropy loss commonly used for segmentation. As in this case there are only two class categories, the loss is the following: whereŜ t is the predicted segmentation and S t is the ground truth segmentation using a binary encoding S t = {0, 1}.

Referring expression categorization
We propose a novel categorization for referring expressions (REs), i.e., linguistic expressions that allow the identification of an individual object (the referent) in a discourse or scene. Our categorization is adapted to the challenges posed by the video object segmentation (VOS) task. We follow the commonly adopted definition of REs put forward by computational linguistics and natural language processing (e.g., [36]), and consider a (noun) phrase as a RE if it is an accurate description of the referent, but not of any other object in the current scene. Likewise, in the vision & language research field, visual RE resolution and generation has seen a rise of interest, especially in still images [8,28,30,31,50], and more recently also on videos [1,6]. The task is formulated as, given an instance comprising an image or video with one or multiple objects, and a RE, identify the referent that the RE describes by predicting, e.g., its bounding box or segmentation mask. The difficulty of the task increases with the number of objects appearing in the scene, and the number of objects of the same class (e.g., multiple dogs). Such cases require more complex REs in order to identify the referent.
In order to make progress on VOS with REs and allow for a systematic comparison of methods, benchmark datasets need to be challenging from both the visual and linguistic perspective. However, for example, most video sequences in the DAVIS-2017 dataset used in Khoreva et al. [21] show a single object in the scene or, at most, different objects from different classes. In these cases, the actual challenge is that of predicting accurate object masks for the RE. On the other hand, the existing datasets for VOS with REs do not focus on the particularities that video information provides either, and often use object attributes which can be already captured by a single frame, or are not even true for the whole clip (e.g., the A2D dataset provides linguistic expressions for only a few frames per clip).
Our novel categorization of REs for video objects allows the analysis of datasets with respect to the difficulty of the REs and the kind of semantic information they provide. We apply it to label and analyze existing expressions of DAVIS-2017 and A2D. In addition, we use this categorization to extend a subset of the A2D test set with REs which contain semantically varying information to analyze how our model behaves with respect to the different categories.

Difficulty and correctness of datasets
We first assess the validity and visual difficulty of a subset of DAVIS-2017 and A2D, by classifying each instance (an object and its linguistic expression) into trivial or non-trivial: if the referent is not the only object of a certain class in the video (e.g., multiple dogs) we consider it non-trivial, otherwise trivial. We further label each expression according to its linguistic ambiguity and correctness: we mark it as no RE if its referent is not the only object in the video which could be described by the expression, i.e., if it does not comply with our definition of a RE given above, and as wrong object if it does not match the target object of the instance but another object in the same video.

Data and Annotation Procedure
Annotation was performed on the DAVIS-2017 validation set (61 REs provided by annotator 1 [21]) in the full video setup (see Section 2), as well as on the subset of the A2D test set which contains at least two annotated objects (856 instances). Each instance contained therein was annotated by one out of four persons (all co-authors). Note that we assume the instances in A2D videos with only a single annotation as trivial, and automatically labeled them as such (439 instances). Fig. 3 shows the proportion of expressions in the DAVIS-2017 and A2D sets with respect to their difficulty, ambiguity, and correctness. Despite being collected in a (non-interactive) referential two-player game setup, DAVIS-2017 contains a considerable proportion of ambiguous expressions (no RE, 8%). The proportion in A2D is slighlty higher (11%), but note that A2D was designed to contain descriptive expressions in contrast to unique references (as defined above). About 52% in DAVIS and 35% in A2D are nontrivial expressions, that is, more challenging for LVOS from both the linguistic and visual perspective, since the object class itself is not sufficient to uniquely identify the target object.

Semantic categorization of REs
Our categorization, shown in Table 1, is inspired by semantic categories of situations and utterances in linguistics [13,23], tailored to the situations found in video data. Specifically, we analyze the REs with respect to the type of information they express, by assigning them categories assumed to be relevant for reference to objects in visual scenes. We focus on information relevant for both, objects in still images and videos, namely the category, appearance, and the location of the referent, and distinguish between information assumed to be more relevant for videos only, namely motion vs. static events. If, according to the RE, the referent acts upon other objects in the scene, we distinguish between whether an object is moved by the referent or not (obj-motion vs. obj-static). This information may be particularly valuable for models that reason over object interactions.
(Psycho)linguistic studies have observed a tendency of REs to contain redundant nondiscriminating information, i.e., logically more information than required to establish unique reference, arguably because this reduces the effort needed for identification [15,23]. In particular the kind (category) of the object and salient properties, such as color, have been found to be used redundantly [37]. To assess whether the phenomenon of redundancy is born out in the video datasets, we additionally label instances as redundant or minimal. A RE is labeled as minimal if it does not include more information than required to identify the target object, and redundant otherwise.

Data and annotation procedure
We collect annotations for the same 61 instances of the validation set of DAVIS-2017 as above, and for a subset of the test set of A2D, which we call A2Dre henceforth. We obtain A2Dre by selecting only instances that were labeled as non-trivial, which are 433 REs from 190 videos. We do not use the trivial cases since the analysis of such examples is not relevant, as referents can be described by using the category alone. Each annotator was presented with a RE, a video in which the target object was marked by a bounding box, and a set of questions paraphrasing our categories (see Table 1). Three annotators (all co-authors of the paper) individually labeled all instances of the DAVIS-2017 val set. The inter-annotator agreement can be considered substantial for all categories, with Davies & Fleiss' kappa coefficients [9] between κ = .83 and .97 (except obj-static, κ = .35, which has only 5 positively labeled instances by at most 2 annotators, and category which obtained perfect agreement). A2Dre was subsequently annotated by the same 3 annotators. Our final set of category annotations used for analysis was derived by means of majority voting: for each non-trivial RE, we kept all category labels which were assigned to the RE by at least two annotators.

Results: What kind of information do REs express?
First of all, we found 99% of the REs for non-trivial instances in A2Dre, and 66% in DAVIS-2017 val (74% including trivial), respectively, to contain redundant information. Recall that only the REs in DAVIS-2017 were obtained in a referential setup, the relatively larger proportion of redundant REs in A2D is therefore not surprising. Figure 4 shows the proportion of instances in the two datasets (DAVIS-2017 val and A2Dre) that were labeled with the individual categories. As expected, the name or category of the referent is virtually always expressed. The visual properties of the referent, i.e., appearance, is prominent in both datasets, too (approx. 60%). Taken together with their high redundancy ratio, this confirms what has been found in psycholinguistic studies on reference [23]. The remaining categories, however, are rare in both datasets, or are only highly frequent in A2Dre, with location and motion being used in the majority of REs. That A2Dre comprises more complex REs than DAVIS-2017 may be not only due to their collection as descriptive, instead of discrimininative phrases, but also due to the much higher complexity of the video scenes. Note that information about referent-object interactions (obj-static and obj-motion) is neglectable, which illustrates the datasets' limited usefulness for research on  [41,45,51]. In the experiments we report in Section 5, we discard these categories, and focus on the remaining categories only, for which we augment the A2Dre dataset.

Extending A2D with REs
As explained above, A2Dre is a subset from the A2D test set including 433 non-trivial REs. Due to its highly unbalanced distribution across the 7 semantic categories (Fig. 4), we select the 4 major categories appearance, location, motion and static. The four categories have in common that in most cases, for a given referent, a RE can be provided that expresses a certain category, and one that does not. We use these categories to augment A2Dre with additional REs, which vary according to the presence or absence of each of them. Specifically, based on our categorization of the original REs, for each RE re and category C, we produce an additional RE re by modifying re slightly such that it does (or does not) express C. For example, for the last RE in Fig. 6, i.e. girl in yellow dress standing near the woman, which could be categorized as appearance (App+), location (Loc+), no motion (Motion-) and static (Static+), we produce new REs for each category: girl standing near the woman (no appearance, App-), girl in yellow dress standing (no location, Loc-), girl in yellow dress walking (motion, Motion+) and girl in yellow dress near the woman (no static, Static-). We do not apply this procedure for category, since it is expressed in almost all REs, and its removal may be difficult in many cases. We will refer to this extended dataset as A2Dre+.

Experiments
We report results with our model on two different tasks: language-guided image segmentation and language-guided video object segmentation (LVOS). The results for still images are obtained on RefCOCO and RefCOCO+ [49], while those for videos correspond to DAVIS-2017 and A2D. Table 2 shows the impact of BERT embeddings in our model on both RefCOCO and RefCOCO+, compared with a bidirectional LSTM similar to Chen et al. [4] for encoding the linguistic expression. In particular, we average the GloVe embeddings [32] of each MLM loss refers to the masked language modelling loss used training BERT with the REs from RefCOCO token and concatenate the mean embeddings of the forward and backward pass. This baseline is compared to two configurations that use BERT. The first one fine-tunes BERT for the language-guided image segmentation task, and significantly boosts performance over using GloVe embeddings. The second has an additional step, that consists in first training BERT using the masked language modelling (MLM) loss with the REs from RefCOCO, as explained in Section 3, and then fine-tuning BERT on the language-guided image segmentation task (as in the previous configuration). We see that this configuration brings an additional gain. Table 2 also compares our model with the state of the art on language-guided image segmentation. STEP [4] consists of an iterative model that refines the RE representation to improve the segmentation. Note that the model must be run for each iteration. Our model surpasses STEP (1-fold), which corresponds to a comparable computational cost, on Ref-COCO val and testA, and is still slightly better than STEP (4-fold). Compared to STEP (5-fold), the performance of our method is slightly lower. BRINet [17] and CMPC [18] are both superior in terms of performance. However, compared to ours, they are significantly more complex. CMPC is composed of several independent modules and needs to build a relational graph per query. BRINet has a cross-attentional and a bidirectional module to fuse cross-modal features. Both BRINet and CPMC use a Dense-CRF post-processing step [22]. In comparison, our network is simpler and is fully end-to-end trainable. Qualitative results generated with our best model on RefCOCO are depicted in Fig. 5. We note how our model distinguishes properly the referred instance and generates an accurate mask. We conclude that our approach is competitive with the state of the art for language-guided image segmentation. Hence, RefVOS is a valid model for language-guided VOS, and for running an analysis on our RE categorization.

Language-guided video object segmentation (LVOS)
Our model is assessed for LVOS on DAVIS-2017 and A2D. In both cases, each video frame is treated separately, so we use the same architecture as in the experiments on still images in Section 5.1.
Our experiments on the DAVIS-2017 validation set are reported in Table 3. All models are pre-trained on RefCOCO. Results are provided with the J&F metric adopted in the DAVIS-2017 challenge for the two different types of REs collected by Khoreva et al. [21] explained in Section 2. J&F is the average between a region-based evaluation measure (J) and a contour-based one (F). Our experiments indicate that our baseline model trained only with RefCOCO already outperforms the best model by Khoreva et al. [21], despite the latter being fine-tuned on the same DAVIS-2017 dataset (+Ft DAVIS segms.). The difference increases when our model is fine-tuned with the segmentations provided in the training set, but freezing the language encoder. This is the configuration comparable to Khoreva et al. [21] in terms of training data, and brings gains of 2.7 and 4.9 points for the first frame and full video REs, respectively. Finally, we also fine-tune the BERT language encoder, obtaining a significant extra gain in performance. We want to highlight that our frame-based model does not rely on any post-processing to add temporal coherence, or optical flow, in contrast to Khoreva et al. [21], so our method may be more efficient computationally.
We also compare our model to URVOS [38], a concurrent work to ours. RefVOS performs slightly better when trained with the same amount of annotated data. Qualitative results for full video REs are shown in Fig. 1. When the multiple objects belong to different categories, the model produces accurate masks from the language query, whereas it is more challenging to properly segment the referent in cases where there are multiple instances of the same class in the sequence (3rd row). The fine-tuning is done with the full video REs, and the REs shown in Fig. 1 are of the same kind. We note how the referred object is in general identified and properly segmented. The results for A2D are shown in Table 4, using the metrics that allow us a comparison with previous works [12,40]. Our model trained only with A2D already outperforms Gavrilyuk et al. [12] in Precision at a high threshold and at the Overall and Mean Intersection Over Union (IoU). Moreover, our model significantly increases its performance when it is first trained on RefCOCO and later fine-tuned on A2D, both its visual and language branches. In this setup, it achieves state of the art results in all metrics by significant margins. Note that both Gavrilyuk et al. [12] and Wang et al. [40] leverage an encoder that was pre-trained on the Kinetics dataset, which includes 650,000 video clips [3]. Hence, these models see a large amount of annotated data for action recognition in videos. We also want to stress our higher Precision values at high thresholds (Prec@0.9) compared to previous works [12,40], which indicate that our model is able to produce more accurate masks. Visualizations with our model are illustrated in Fig. 6.
In conclusion, RefVOS is state of the art for DAVIS-2017 and A2D on the LVOS task, although it is a frame-based model. This motivates the analysis of our model when tested with different types of REs, based on the categorization and difficulty analysis proposed in Section 4.

Referring expressions analysis for LVOS
Firstly, we analyze the performance on trivial and non-trivial linguistic expressions for both the A2D test and DAVIS-2017 validation sets. The mean IoU per referent obtained for trivial and non-trivial is 48.7 vs. 46.2 on DAVIS-2017, and 53.9 vs. 33.2 on A2D. We observe that the performance is worse for the non-trivial cases for both datasets as expected, with a major drop on A2D.

Fig. 6
Video sequences for A2D with expressions (language queries) and the results of our model. The first column shows a reference frame, the second to fourth columns depict the masks produced by our model when given the expression shown on top. Finally, the fifth to seventh columns show the results for the expression shown on top of these columns, which refers to another object of the video sequence Secondly, we study the effect of RE categories in relation to the performance of RefVOS. The A2Dre+ dataset described in Section 4.3 allows us to have the same number of referents for all major categories: appearance, location, motion and static. Each of our referents is annotated with highly similar REs (two for each category) and are hence directly comparable. In contrast, Khoreva et al. [21] split the videos into two different subsets with non-comparable referents. Table 5 compares the performance of RefVOS depending on whether each of the categories C is present in the RE (C+, the RE expresses C) or not (C-). The results show that the presence of appearance (App+) and location (Loc+) categories yields significantly higher results compared to their absence (App-and Loc-, respectively). We also observe a drop in performance when the static category is present (Static+), which indicates that the model struggles at identifying a referent based on static actions such as holding, sitting, eating. In contrast, the presence (Motion+) or absence (Motion-) of the motion category does not affect the performance, which actually means that the model is unable to benefit from this type of REs.
In what follows, we further visually analyze the results obtained with RefVOS depending on the categories that appear in the corresponding REs. Figure 7 includes examples of the results of our model with A2Dre+. We analyze the categories appearance, location, motion and static, which are the most common semantic categories. Each column is a first frame of a video sequence with a non-trivial case, and each row is a different RE that has or has not a certain category. As we concluded with the quantitative results, the performance  Fig. 7 Each column is the first frame of a video sequence of A2D. Each row indicates if the RE that produces the depicted result does (C+) or does not (C-) contain a certain category C from our proposed categorization. The framed REs are the original REs in the dataset [12]. For the example in the fourth column, the natural expression in the annotations from Gavrilyuk et al. [12] is man standing, which is not a RE as it does not uniquely identify the target object as defined in Section 4. For this reason, for this example all REs shown in the Figure are our own annotations when the appearance and location categories are present is higher compared to when these categories are absent. Regarding the motion and static categories, we first notice that the annotators considered it impossible for some cases to create corresponding REs. We indicate those examples with the "Not Applicable" label. We see how the presence or the absence of the motion and static categories has a minimal impact to the results. In fact, adding these categories (through a description of the shown action, e.g., "crawling" for Motion+ or "eating" for Static+) even leads to worse segmentations, like in the example "a man eating a big sandwich" from the second column in Fig. 7.
Finally, in Table 6 we study the effect of feeding the model with only the actor, the action, or the actor and action, without formulating any RE, for all the test set of A2D. These actor and action terms are obtained from the dataset collected by Gavrilyuk et al. [12]. In most cases these expressions are not REs as they do not unambiguously describe the referent in the video (cf. Section 4). Additionally, we consider a generic phrase thing. We distinguish between trivial and non-trivial cases. Results show that RefVOS works significantly better when the actor is provided than when the action is. Furthermore, performance improves when using both. Finally, having the full linguistic phrase is still the best model. Remarkably, our configuration with actor and action reaches higher Overall IoU than previous works that use complete linguistic phrases (see Table 4). Note that using the full phrase improves performance especially for the non-trivial cases, as these require complete linguistic expressions to identify the referent. We also want to stress that the aggregated performance, i.e., considering all cases, is dominated by the performance of the trivial ones, as they represent most of the dataset.

Conclusions
This work studies the difficulty of REs from benchmarks on Language-guided Video Object Segmentation (LVOS), and proposes seven semantic categories to analyze the nature of such REs. We introduce RefVOS, a novel model that is competitive for language-guided image segmentation, and state of the art for LVOS. However, our analysis shows that benchmarks are mainly composed of trivial cases, in which referents can be identified with simple linguistic expressions. This indicates that the reported metrics for the task may be misleading. Thus, we focus on the non-trivial cases. We extend A2D with new REs with diverse semantic categories for non-trivial cases, and test our model with them, which reveals that it struggles at exploiting motion and static events, and that it mainly benefits from REs based on appearance and location. We reckon that future research on LVOS should focus on non-trivial cases describing motion and events, as they present a challenge for language grounding on videos. Concurrent to our work, Seo et al. [38] collected Refer-Youtube-VOS, a large-scale benchmark for LVOS built on top of Youtube-VOS [44]. We believe that, as future work, our categorization for REs could be used to classify the provided linguistic expressions by this benchmark. Thus, models could be evaluated based on the non-trivial cases and the different categories in order to analyze which REs are more challenging when using a large-scale dataset.
Funding Open Access funding provided thanks to the CRUE-CSIC agreement with Springer Nature. This work was partially supported by the projects PID2019-107255GB-C22 and PID2020-117142GB-I00 funded by MCIN/ AEI /10.13039/501100011033 Spanish Ministry of Science, and the grant 2017-SGR-1414 of the Government of Catalonia. This work was also partially supported by the project RTI2018-095232-B-C22 funded by the Spanish Ministry of Science, Innovation and Universities.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.