Referring Expression Segmentation

The RefCOCO dataset is a referring expression generation (REG) dataset used for tasks related to understanding natural language expressions that refer to specific objects in images. Here are the key details about RefCOCO: 1. Collection Method: The dataset was collected using the ReferitGame, a two-player game. In this game, the first player views an image with a segmented target object and writes a natural language expression referring to that object. The second player sees only the image and the referring expression and must click on the corresponding object. If both players perform correctly, they earn points and switch roles; otherwise, they receive a new object and image for description. 2. Dataset Variants: RefCOCO: Contains 142,209 refer expressions for 50,000 objects across 19,994 images. RefCOCO+: Includes 141,564 expressions for 49,856 objects in 19,992 images. RefCOCOg: This variant has 25,799 images, 95,010 referring expressions, and 49,822 object instances. 3. Language and Restrictions: RefCOCO allows any type of language in the referring expressions. RefCOCO+ disallows location words in expressions to focus purely on appearance-based descriptions (e.g., "the man in the yellow polka-dotted shirt") rather than viewer-dependent descriptions (e.g., "the second man from the left"). These datasets serve as valuable resources for tasks like referring expression segmentation, comprehension, and visual grounding in computer vision research.

DAVIS17 is a dataset for video object segmentation. It contains a total of 150 videos - 60 for training, 30 for validation, 60 for testing Source: Siam R-CNN: Visual Tracking by Re-Detection Image Source: https://www.researchgate.net/figure/LucidTracker-qualitative-results-on-DAVIS-17-test-dev-set-Frames-sampled-along-the_fig5_331792902

JHMDB is an action recognition dataset that consists of 960 video sequences belonging to 21 actions. It is a subset of the larger HMDB51 dataset collected from digitized movies and YouTube videos. The dataset contains video and annotation for puppet flow per frame (approximated optimal flow on the person), puppet mask per frame, joint positions per frame, action label per clip and meta label per clip (camera motion, visible body parts, camera viewpoint, number of people, video quality). Source: Unsupervised Deep Metric Learning via Orthogonality based Probabilistic Loss Image Source: https://arxiv.org/pdf/1712.06316.pdf

Our task is to localize and provide a pixel-level mask of an object on all video frames given a language referring expression obtained either by looking at the first frame only or the full video. To validate our approach we employ two popular video object segmentation datasets, DAVIS16 [38] and DAVIS17 [42]. These two datasets introduce various challenges, containing videos with single or multiple salient objects, crowded scenes, similar looking instances, occlusions, camera view changes, fast motion, etc. DAVIS16 [38] consists of 30 training and 20 test videos of diverse object categories with all frames annotated with pixel-level accuracy. Note that in this dataset only a single object is annotated per video. For the multiple object video segmentation task we consider DAVIS17. Compared to DAVIS16, this is a more challenging dataset, with multiple objects annotated per video and more complex scenes with more distractors, occlusions, smaller objects, and fine structures. Overall, DAVIS17 consists of a training set with 60 videos, and a validation/test-dev/test-challenge set with 30 sequences each. As our goal is to segment objects in videos using language specifications, we augment all objects annotated with mask labels in DAVIS16 and DAVIS17 with non-ambiguous referring expressions. We follow the work of [34] and ask the annotator to provide a language description of the object, which has a mask annotation, by looking only at the first frame of the video. Then another annotator is given the first frame and the corresponding description, and asked to identify the referred object. If the annotator is unable to correctly identify the object, the description is corrected to remove ambiguity and to specify the object uniquely. We have collected two referring expressions per target object annotated by non-computer vision experts (Annotator 1, 2). However, by looking only at the 1st frame, the obtained referring expressions may potentially be invalid for an entire video. (We actually quantified that only∼ 15% of the collected descriptions become invalid over time and it does not affect strongly segmentation results as temporal consistency step helps to disambiguate some of such cases, see the supp. material for details.) Besides, in many applications, such as video editing or video-based advertisement, the user has access to a full video. Providing a language query which is valid for all frames might decrease the editing time and result in more coherent predictions. Thus, on DAVIS17 we asked the workers to provide a description of the object by looking at the full video. We have collected one expression of the full video type per target object. Future work may choose to use either setting. The average length for the first frame/full video expressions is 5.5/6.3 words. For DAVIS17 first frame annotations we notice that descriptions given by Annotator 1 are longer than the ones by Annotator 2 (6.4 vs. 4.6 words). We evaluate the effect of description length on the grounding performance in §5. Besides, the expressions relevant to a full video mention verbs more often than the first frame descriptions (44% vs. 25%). This is intuitive, as referring to an object which changes its appearance and position over time may require mentioning its actions. Adjectives are present in over 50% for all annotations. Most of them refer to colors (over 70%), shapes and sizes (7%) and spatial/ordering words (6% first frame vs. 13% full video expressions). The full video expressions also have a higher number of adverbs and prepositions, and overall are more complex than the ones provided for the first frame. Overall augmented DAVIS16/17 contains ∼ 1.2k referring expressions for more than 400 objects on 150 videos with ∼ 10k frames. We believe the collected data will be of interest to segmentation as well as vision and language communities, providing an opportunity to explore language as alternative input for video object segmentation.

There exist previous works [6, 10] that constructed referring segmentation datasets for videos. Gavrilyuk et al. [6] extended the A2D [33] and J-HMDB [9] datasets with natural sentences; the datasets focus on describing the ‘actors’ and ‘actions’ appearing in videos, therefore the instance annotations are limited to only a few object categories corresponding to the dominant ‘actors’ performing a salient ‘action’. Khoreva et al. [10] built a dataset based on DAVIS [25], but the scales are barely sufficient to learn an end-to-end model from scratch Youtube-VOS has 4,519 high-resolution videos with 94 common object categories. Each video has pixel-level instance segmentation annotation at every 5 frames in 30-fps videos, and their durations are around 3 to 6 seconds. We employed Amazon Mechanical Turk to annotate referring expressions. To ensure the quality of the annotations, we selected around 50 turkers after a validation test. Each turker was given a pair of videos, the original video and the mask-overlaid one with the target object highlighted, and was asked to provide a discriminative sentence within 20 words that describes the target object accurately. We collected two kinds of annotations, which describe the highlighted object (1) based on a whole video (Full-video expression) and (2) using only the first frame of the video (First-frame expression). After the initial annotation, we conducted verification and cleaning jobs for all annotations, and dropped objects if an object cannot be localized using language expressions only. The followings are the statistics and analysis of the two annotation types of the dataset after the verification. Full-video expression: Youtube-VOS has 6,459 and 1,063 unique objects in train and validation split, respectively. Among them, we cover 6,388 unique objects in 3,471 videos (6, 388/6, 459 = 98.9%) with 12,913 expressions in train split and 1,063 unique objects in 507 videos (1, 063/1, 063 = 100%) with 2,096 expressions in validation split. On average, each video has 3.8 language expressions and each expression has 10.0 words. First-frame expression: There are 6,006 unique objects in 3,412 videos (6, 006 /6, 459 = 93.0%) with 10,897 expressions in train split and 1,030 unique objects in 507 videos (1, 030/1, 063 = 96.9%) with 1,993 expressions in validation split. The number of annotated objects is lower than that of the full-video expressions because using only the first frame makes annotation more ambiguous and inconsistent and we dropped more annotations during the verification. On average, each video has 3.2 language expressions and each expression has 7.5 words.

A new large-scale dataset for referring expressions, based on MS-COCO. Source: Generation and Comprehension of Unambiguous Object Descriptions

gRefCOCO is the first large-scale Generalized Referring Expression Segmentation dataset that contains multi-target, no-target, and single-target expressions.

The Actor-Action Dataset (A2D) by Xu et al. [29] serves as the largest video dataset for the general actor and action segmentation task. It contains 3,782 videos from YouTube with pixel-level labeled actors and their actions. The dataset includes eight different actions, while a total of seven actor classes are considered to perform those actions. We follow [29], who split the dataset into 3,036 training videos and 746 testing videos. As we are interested in pixel-level actor and action segmentation from sentences, we augment the videos in A2D with natural language descriptions about what each actor is doing in the videos. Following the guidelines set forth in [12], we ask our annotators for a discriminative referring expression of each actor instance if multiple objects are considered in a video. The annotation process resulted in a total of 6,656 sentences, including 811 different nouns, 225 verbs and 189 adjectives. Our sentences enrich the actor and action pairs from the A2D dataset with finer granularities. For example, the actor adult in A2D may be annotated with man, woman, person and player in our sentences, while action rolling may also refer to flipping, sliding, moving and running when describing different actors in different scenarios. Our sentences contain on average more words than the ReferIt dataset [12] (7.3 vs 4.7), even when we leave out prepositions, articles and linking verbs (4.5 vs 3.6). This makes sense as our sentences contain a variety of verbs while existing referring expression datasets mostly ignore verbs.

PhraseCut is a dataset consisting of 77,262 images and 345,486 phrase-region pairs. The dataset is collected on top of the Visual Genome dataset and uses the existing annotations to generate a challenging set of referring phrases for which the corresponding regions are manually annotated. Source: PhraseCut: Language-based Image Segmentation in the Wild Image Source: https://people.cs.umass.edu/~chenyun/publication/phrasecut/

CLEVR-Ref+ is a synthetic diagnostic dataset for referring expression comprehension. The precise locations and attributes of the objects are readily available, and the referring expressions are automatically associated with functional programs. The synthetic nature allows control over dataset bias (through sampling strategy), and the modular programs enable intermediate reasoning ground truth without human annotators. Source: CLEVR-Ref+: Diagnosing Visual Reasoning with Referring Expressions

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 as 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. A2Dre was annotated by 3 authors of the paper. Our final set of category annotations used for analysis was derived by means of majority voting: for each nontrivial RE, we kept all category labels which were assigned to the RE by at least two annotators.

A2Dre is a subset from the A2D test set including $433$\textit{non-trivial} REs. Due to its highly unbalanced distribution across the $7$semantic categories we select the $4$major categories \textsl{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 Figure~\ref{fig:a2d-images}, i.e. \emph{girl in yellow dress standing near the woman}, which could be categorized as \textit{appearance}, \textit{location}, no \textit{motion} and \textit{static}, we produce new REs for each category: \emph{girl standing near the woman} (no \textit{appearance}), \emph{girl in yellow dress standing} (no \textit{location}), \emph{girl in yellow dress walking} (\textit{motion}) and \emph{girl in yellow dress near the woman} (no \textit{static}). We do not apply this procedure for \textsl{category}, since it is expressed in almost all REs, and its removal may be difficult in many cases. We name this extended dataset as A2Dre+.