Rajat Koner

Rajat Koner, Tanveer Hannan, Suprosanna Shit et al. · deepmind

Recent transformer-based offline video instance segmentation (VIS) approaches achieve encouraging results and significantly outperform online approaches. However, their reliance on the whole video and the immense computational complexity caused by full Spatio-temporal attention limit them in real-life applications such as processing lengthy videos. In this paper, we propose a single-stage transformer-based efficient online VIS framework named InstanceFormer, which is especially suitable for long and challenging videos. We propose three novel components to model short-term and long-term dependency and temporal coherence. First, we propagate the representation, location, and semantic information of prior instances to model short-term changes. Second, we propose a novel memory cross-attention in the decoder, which allows the network to look into earlier instances within a certain temporal window. Finally, we employ a temporal contrastive loss to impose coherence in the representation of an instance across all frames. Memory attention and temporal coherence are particularly beneficial to long-range dependency modeling, including challenging scenarios like occlusion. The proposed InstanceFormer outperforms previous online benchmark methods by a large margin across multiple datasets. Most importantly, InstanceFormer surpasses offline approaches for challenging and long datasets such as YouTube-VIS-2021 and OVIS. Code is available at https://github.com/rajatkoner08/InstanceFormer.

Suprosanna Shit, Rajat Koner, Bastian Wittmann et al. · deepmind

A comprehensive representation of an image requires understanding objects and their mutual relationship, especially in image-to-graph generation, e.g., road network extraction, blood-vessel network extraction, or scene graph generation. Traditionally, image-to-graph generation is addressed with a two-stage approach consisting of object detection followed by a separate relation prediction, which prevents simultaneous object-relation interaction. This work proposes a unified one-stage transformer-based framework, namely Relationformer, that jointly predicts objects and their relations. We leverage direct set-based object prediction and incorporate the interaction among the objects to learn an object-relation representation jointly. In addition to existing [obj]-tokens, we propose a novel learnable token, namely [rln]-token. Together with [obj]-tokens, [rln]-token exploits local and global semantic reasoning in an image through a series of mutual associations. In combination with the pair-wise [obj]-token, the [rln]-token contributes to a computationally efficient relation prediction. We achieve state-of-the-art performance on multiple, diverse and multi-domain datasets that demonstrate our approach's effectiveness and generalizability.

Hang Li, Jindong Gu, Rajat Koner et al. · deepmind, oxford

The field of multimodal research focusing on the comprehension and creation of both images and text has witnessed significant strides. This progress is exemplified by the emergence of sophisticated models dedicated to image captioning at scale, such as the notable Flamingo model and text-to-image generative models, with DALL-E serving as a prominent example. An interesting question worth exploring in this domain is whether Flamingo and DALL-E understand each other. To study this question, we propose a reconstruction task where Flamingo generates a description for a given image and DALL-E uses this description as input to synthesize a new image. We argue that these models understand each other if the generated image is similar to the given image. Specifically, we study the relationship between the quality of the image reconstruction and that of the text generation. We find that an optimal description of an image is one that gives rise to a generated image similar to the original one. The finding motivates us to propose a unified framework to finetune the text-to-image and image-to-text models. Concretely, the reconstruction part forms a regularization loss to guide the tuning of the models. Extensive experiments on multiple datasets with different image captioning and image generation models validate our findings and demonstrate the effectiveness of our proposed unified framework. As DALL-E and Flamingo are not publicly available, we use Stable Diffusion and BLIP in the remaining work. Project website: https://dalleflamingo.github.io.

Poulami Sinhamahapatra, Rajat Koner, Karsten Roscher et al.

It is essential for safety-critical applications of deep neural networks to determine when new inputs are significantly different from the training distribution. In this paper, we explore this out-of-distribution (OOD) detection problem for image classification using clusters of semantically similar embeddings of the training data and exploit the differences in distance relationships to these clusters between in- and out-of-distribution data. We study the structure and separation of clusters in the embedding space and find that supervised contrastive learning leads to well-separated clusters while its self-supervised counterpart fails to do so. In our extensive analysis of different training methods, clustering strategies, distance metrics, and thresholding approaches, we observe that there is no clear winner. The optimal approach depends on the model architecture and selected datasets for in- and out-of-distribution. While we could reproduce the outstanding results for contrastive training on CIFAR-10 as in-distribution data, we find standard cross-entropy paired with cosine similarity outperforms all contrastive training methods when training on CIFAR-100 instead. Cross-entropy provides competitive results as compared to expensive contrastive training methods.

Rajat Koner, Gagan Jain, Prateek Jain et al.

Vision Transformers (ViT) have emerged as the de-facto choice for numerous industry grade vision solutions. But their inference cost can be prohibitive for many settings, as they compute self-attention in each layer which suffers from quadratic computational complexity in the number of tokens. On the other hand, spatial information in images and spatio-temporal information in videos is usually sparse and redundant. In this work, we introduce LookupViT, that aims to exploit this information sparsity to reduce ViT inference cost. LookupViT provides a novel general purpose vision transformer block that operates by compressing information from higher resolution tokens to a fixed number of tokens. These few compressed tokens undergo meticulous processing, while the higher-resolution tokens are passed through computationally cheaper layers. Information sharing between these two token sets is enabled through a bidirectional cross-attention mechanism. The approach offers multiple advantages - (a) easy to implement on standard ML accelerators (GPUs/TPUs) via standard high-level operators, (b) applicable to standard ViT and its variants, thus generalizes to various tasks, (c) can handle different tokenization and attention approaches. LookupViT also offers flexibility for the compressed tokens, enabling performance-computation trade-offs in a single trained model. We show LookupViT's effectiveness on multiple domains - (a) for image-classification (ImageNet-1K and ImageNet-21K), (b) video classification (Kinetics400 and Something-Something V2), (c) image captioning (COCO-Captions) with a frozen encoder. LookupViT provides $2\times$ reduction in FLOPs while upholding or improving accuracy across these domains. In addition, LookupViT also demonstrates out-of-the-box robustness and generalization on image classification (ImageNet-C,R,A,O), improving by up to $4\%$ over ViT.

Tanveer Hannan, Rajat Koner, Maximilian Bernhard et al.

Recent trends in Video Instance Segmentation (VIS) have seen a growing reliance on online methods to model complex and lengthy video sequences. However, the degradation of representation and noise accumulation of the online methods, especially during occlusion and abrupt changes, pose substantial challenges. Transformer-based query propagation provides promising directions at the cost of quadratic memory attention. However, they are susceptible to the degradation of instance features due to the above-mentioned challenges and suffer from cascading effects. The detection and rectification of such errors remain largely underexplored. To this end, we introduce \textbf{GRAtt-VIS}, \textbf{G}ated \textbf{R}esidual \textbf{Att}ention for \textbf{V}ideo \textbf{I}nstance \textbf{S}egmentation. Firstly, we leverage a Gumbel-Softmax-based gate to detect possible errors in the current frame. Next, based on the gate activation, we rectify degraded features from its past representation. Such a residual configuration alleviates the need for dedicated memory and provides a continuous stream of relevant instance features. Secondly, we propose a novel inter-instance interaction using gate activation as a mask for self-attention. This masking strategy dynamically restricts the unrepresentative instance queries in the self-attention and preserves vital information for long-term tracking. We refer to this novel combination of Gated Residual Connection and Masked Self-Attention as \textbf{GRAtt} block, which can easily be integrated into the existing propagation-based framework. Further, GRAtt blocks significantly reduce the attention overhead and simplify dynamic temporal modeling. GRAtt-VIS achieves state-of-the-art performance on YouTube-VIS and the highly challenging OVIS dataset, significantly improving over previous methods. Code is available at \url{https://github.com/Tanveer81/GRAttVIS}.

Tanveer Hannan, Shuaicong Wu, Mark Weber et al.

Understanding fine-grained actions and accurately localizing their corresponding actors in space and time are fundamental capabilities for advancing next-generation AI systems, including embodied agents, autonomous platforms, and human-AI interaction frameworks. Despite recent progress in video understanding, existing methods predominantly address either coarse-grained action recognition or generic object tracking, thereby overlooking the challenge of jointly detecting and tracking multiple objects according to their actions while grounding them temporally. To address this gap, we introduce Spatio-temporal Video Action Grounding (SVAG), a novel task that requires models to simultaneously detect, track, and temporally localize all referent objects in videos based on natural language descriptions of their actions. To support this task, we construct SVAG-Bench, a large-scale benchmark comprising 688 videos, 19,590 annotated records, and 903 unique verbs, covering a diverse range of objects, actions, and real-world scenes. We further propose SVAGFormer, a baseline framework that adapts state of the art vision language models for joint spatial and temporal grounding, and introduce SVAGEval, a standardized evaluation toolkit for fair and reproducible benchmarking. Empirical results show that existing models perform poorly on SVAG, particularly in dense or complex scenes, underscoring the need for more advanced reasoning over fine-grained object-action interactions in long videos.

Roberto Amoroso, Gengyuan Zhang, Rajat Koner et al.

Video Question Answering (Video QA) is a challenging video understanding task that requires models to comprehend entire videos, identify the most relevant information based on contextual cues from a given question, and reason accurately to provide answers. Recent advancements in Multimodal Large Language Models (MLLMs) have transformed video QA by leveraging their exceptional commonsense reasoning capabilities. This progress is largely driven by the effective alignment between visual data and the language space of MLLMs. However, for video QA, an additional space-time alignment poses a considerable challenge for extracting question-relevant information across frames. In this work, we investigate diverse temporal modeling techniques to integrate with MLLMs, aiming to achieve question-guided temporal modeling that leverages pre-trained visual and textual alignment in MLLMs. We propose T-Former, a novel temporal modeling method that creates a question-guided temporal bridge between frame-wise visual perception and the reasoning capabilities of LLMs. Our evaluation across multiple video QA benchmarks demonstrates that T-Former competes favorably with existing temporal modeling approaches and aligns with recent advancements in video QA.

Wenhui Zhu, Xiwen Chen, Zhipeng Wang et al.

Instructed Visual Segmentation (IVS) tasks require segmenting objects in images or videos based on natural language instructions. While recent multimodal large language models (MLLMs) have achieved strong performance on IVS, their inference cost remains a major bottleneck, particularly in video. We empirically analyze visual token sampling in MLLMs and observe a strong correlation between subset token coverage and segmentation performance. This motivates our design of a simple and effective token pruning method that selects a compact yet spatially representative subset of tokens to accelerate inference. In this paper, we introduce a novel visual token pruning method for IVS, called EVTP-IV, which builds upon the k-center by integrating spatial information to ensure better coverage. We further provide an information-theoretic analysis to support our design. Experiments on standard IVS benchmarks show that our method achieves up to 5X speed-up on video tasks and 3.5X on image tasks, while maintaining comparable accuracy using only 20% of the tokens. Our method also consistently outperforms state-of-the-art pruning baselines under varying pruning ratios.

Rajat Koner, Zhipeng Wang, Srinivas Parthasarathy et al.

Online video segmentation methods excel at handling long sequences and capturing gradual changes, making them ideal for real-world applications. However, achieving temporally consistent predictions remains a challenge, especially with gradual accumulation of noise or drift in on-line propagation, abrupt occlusions and scene transitions. This paper introduces Local2Global, an online framework, for video instance segmentation, exhibiting state-of-the-art performance with simple baseline and training purely in online fashion. Leveraging the DETR-based query propagation framework, we introduce two novel sets of queries:(1) local queries that capture initial object-specific spatial features from each frame and (2) global queries containing past spatio-temporal representations. We propose the L2G-aligner, a novel lightweight transformer decoder, to facilitate an early alignment between local and global queries. This alignment allows our model to effectively utilize current frame information while maintaining temporal consistency, producing a smooth transition between frames. Furthermore, L2G-aligner is integrated within the segmentation model, without relying on additional complex heuristics, or memory mechanisms. Extensive experiments across various challenging VIS and VPS datasets showcase the superiority of our method with simple online training, surpassing current benchmarks without bells and rings. For instance, we achieve 54.3 and 49.4 AP on Youtube-VIS-19/-21 datasets and 37.0 AP on OVIS dataset respectively withthe ResNet-50 backbone.

Tanveer Hannan, Rajat Koner, Jonathan Kobold et al.

Video Object Segmentation (VOS) has been targeted by various fully-supervised and self-supervised approaches. While fully-supervised methods demonstrate excellent results, self-supervised ones, which do not use pixel-level ground truth, attract much attention. However, self-supervised approaches pose a significant performance gap. Box-level annotations provide a balanced compromise between labeling effort and result quality for image segmentation but have not been exploited for the video domain. In this work, we propose a box-supervised video object segmentation proposal network, which takes advantage of intrinsic video properties. Our method incorporates object motion in the following way: first, motion is computed using a bidirectional temporal difference and a novel bounding box-guided motion compensation. Second, we introduce a novel motion-aware affinity loss that encourages the network to predict positive pixel pairs if they share similar motion and color. The proposed method outperforms the state-of-the-art self-supervised benchmark by 16.4% and 6.9% $\mathcal{J}$ &$\mathcal{F}$ score and the majority of fully supervised methods on the DAVIS and Youtube-VOS dataset without imposing network architectural specifications. We provide extensive tests and ablations on the datasets, demonstrating the robustness of our method.

Rajat Koner, Poulami Sinhamahapatra, Karsten Roscher et al.

A serious problem in image classification is that a trained model might perform well for input data that originates from the same distribution as the data available for model training, but performs much worse for out-of-distribution (OOD) samples. In real-world safety-critical applications, in particular, it is important to be aware if a new data point is OOD. To date, OOD detection is typically addressed using either confidence scores, auto-encoder based reconstruction, or by contrastive learning. However, the global image context has not yet been explored to discriminate the non-local objectness between in-distribution and OOD samples. This paper proposes a first-of-its-kind OOD detection architecture named OODformer that leverages the contextualization capabilities of the transformer. Incorporating the trans\-former as the principal feature extractor allows us to exploit the object concepts and their discriminate attributes along with their co-occurrence via visual attention. Using the contextualised embedding, we demonstrate OOD detection using both class-conditioned latent space similarity and a network confidence score. Our approach shows improved generalizability across various datasets. We have achieved a new state-of-the-art result on CIFAR-10/-100 and ImageNet30.

Rajat Koner, Hang Li, Marcel Hildebrandt et al.

Visual Question Answering (VQA) is concerned with answering free-form questions about an image. Since it requires a deep semantic and linguistic understanding of the question and the ability to associate it with various objects that are present in the image, it is an ambitious task and requires multi-modal reasoning from both computer vision and natural language processing. We propose Graphhopper, a novel method that approaches the task by integrating knowledge graph reasoning, computer vision, and natural language processing techniques. Concretely, our method is based on performing context-driven, sequential reasoning based on the scene entities and their semantic and spatial relationships. As a first step, we derive a scene graph that describes the objects in the image, as well as their attributes and their mutual relationships. Subsequently, a reinforcement learning agent is trained to autonomously navigate in a multi-hop manner over the extracted scene graph to generate reasoning paths, which are the basis for deriving answers. We conduct an experimental study on the challenging dataset GQA, based on both manually curated and automatically generated scene graphs. Our results show that we keep up with a human performance on manually curated scene graphs. Moreover, we find that Graphhopper outperforms another state-of-the-art scene graph reasoning model on both manually curated and automatically generated scene graphs by a significant margin.

Rajat Koner, Poulami Sinhamahapatra, Volker Tresp

Identifying objects in an image and their mutual relationships as a scene graph leads to a deep understanding of image content. Despite the recent advancement in deep learning, the detection and labeling of visual object relationships remain a challenging task. This work proposes a novel local-context aware architecture named relation transformer, which exploits complex global objects to object and object to edge (relation) interactions. Our hierarchical multi-head attention-based approach efficiently captures contextual dependencies between objects and predicts their relationships. In comparison to state-of-the-art approaches, we have achieved an overall mean \textbf{4.85\%} improvement and a new benchmark across all the scene graph generation tasks on the Visual Genome dataset.

Marcel Hildebrandt, Hang Li, Rajat Koner et al.

Visual question answering is concerned with answering free-form questions about an image. Since it requires a deep linguistic understanding of the question and the ability to associate it with various objects that are present in the image, it is an ambitious task and requires techniques from both computer vision and natural language processing. We propose a novel method that approaches the task by performing context-driven, sequential reasoning based on the objects and their semantic and spatial relationships present in the scene. As a first step, we derive a scene graph which describes the objects in the image, as well as their attributes and their mutual relationships. A reinforcement agent then learns to autonomously navigate over the extracted scene graph to generate paths, which are then the basis for deriving answers. We conduct a first experimental study on the challenging GQA dataset with manually curated scene graphs, where our method almost reaches the level of human performance.

Rajat Koner, Suprosanna Shit, Volker Tresp

The extraction of a scene graph with objects as nodes and mutual relationships as edges is the basis for a deep understanding of image content. Despite recent advances, such as message passing and joint classification, the detection of visual relationships remains a challenging task due to sub-optimal exploration of the mutual interaction among the visual objects. In this work, we propose a novel transformer formulation for scene graph generation and relation prediction. We leverage the encoder-decoder architecture of the transformer for rich feature embedding of nodes and edges. Specifically, we model the node-to-node interaction with the self-attention of the transformer encoder and the edge-to-node interaction with the cross-attention of the transformer decoder. Further, we introduce a novel positional embedding suitable to handle edges in the decoder. Finally, our relation prediction module classifies the directed relation from the learned node and edge embedding. We name this architecture as Relation Transformer Network (RTN). On the Visual Genome and GQA dataset, we have achieved an overall mean of 4.85% and 3.1% point improvement in comparison with state-of-the-art methods. Our experiments show that Relation Transformer can efficiently model context across various datasets with small, medium, and large-scale relation classification.

Sahand Sharifzadeh, Sina Moayed Baharlou, Max Berrendorf et al.

Visual relation detection methods rely on object information extracted from RGB images such as 2D bounding boxes, feature maps, and predicted class probabilities. We argue that depth maps can additionally provide valuable information on object relations, e.g. helping to detect not only spatial relations, such as standing behind, but also non-spatial relations, such as holding. In this work, we study the effect of using different object features with a focus on depth maps. To enable this study, we release a new synthetic dataset of depth maps, VG-Depth, as an extension to Visual Genome (VG). We also note that given the highly imbalanced distribution of relations in VG, typical evaluation metrics for visual relation detection cannot reveal improvements of under-represented relations. To address this problem, we propose using an additional metric, calling it Macro Recall@K, and demonstrate its remarkable performance on VG. Finally, our experiments confirm that by effective utilization of depth maps within a simple, yet competitive framework, the performance of visual relation detection can be improved by a margin of up to 8%.