In-context learning (ICL) enables Large Vision-Language Models (LVLMs) to adapt to new tasks without parameter updates, using a few demonstrations from a large support set. However, selecting informative demonstrations leads to high computational and memory costs. While some methods explore selecting a small and representative coreset in the text classification, evaluating all support set samples remains costly, and discarded samples lead to unnecessary information loss. These methods may also be less effective for image classification due to differences in feature spaces. Given these limitations, we propose Key-based Coreset Optimization (KeCO), a novel framework that leverages untapped data to construct a compact and informative coreset. We introduce visual features as keys within the coreset, which serve as the anchor for identifying samples to be updated through different selection strategies. By leveraging untapped samples from the support set, we update the keys of selected coreset samples, enabling the randomly initialized coreset to evolve into a more informative coreset under low computational cost. Through extensive experiments on coarse-grained and fine-grained image classification benchmarks, we demonstrate that KeCO effectively enhances ICL performance for image classification task, achieving an average improvement of more than 20%. Notably, we evaluate KeCO under a simulated online scenario, and the strong performance in this scenario highlights the practical value of our framework for resource-constrained real-world scenarios.

Foundation models like CLIP allow zero-shot transfer on various tasks without additional training data. Yet, the zero-shot performance is less competitive than a fully supervised one. Thus, to enhance the performance, fine-tuning and ensembling are also commonly adopted to better fit the downstream tasks. However, we argue that such prior work has overlooked the inherent biases in foundation models. Due to the highly imbalanced Web-scale training set, these foundation models are inevitably skewed toward frequent semantics, and thus the subsequent fine-tuning or ensembling is still biased. In this study, we systematically examine the biases in foundation models and demonstrate the efficacy of our proposed Generalized Logit Adjustment (GLA) method. Note that bias estimation in foundation models is challenging, as most pre-train data cannot be explicitly assessed like in traditional long-tailed classification tasks. To this end, GLA has an optimization-based bias estimation approach for debiasing foundation models. As our work resolves a fundamental flaw in the pre-training, the proposed GLA demonstrates significant improvements across a diverse range of tasks: it achieves 1.5 pp accuracy gains on ImageNet, an large average improvement (1.4-4.6 pp) on 11 few-shot datasets, 2.4 pp gains on long-tailed classification. Codes are in https://github.com/BeierZhu/GLA.

Scene graphs connect individual objects with visual relationships. They serve as a comprehensive scene representation for downstream multimodal tasks. However, by exploring recent progress in Scene Graph Generation (SGG), we find that the performance of recent works is highly limited by the pairwise relationship modeling by naive feature concatenation. Such pairwise features lack sufficient object interaction due to the mis-aligned object parts, resulting in non-discriminative pairwise features for visual relationship prediction. For example, naive concatenated pairwise feature usually make the model fail to discriminate betweenridingandfeedingfor object pairpersonandhorse. To this end, we design a meta-architecture— learning-to-align — for dynamic object feature concatenation. We call our model:Align R-CNN. Specifically, we introduce a novel attention-based multiple region alignment module that can be jointly optimized with SGG. Experiments on the large-scale SGG benchmark Visual Genome show that the proposedAlign R-CNNcan replace the naive feature concatenation and thus boost all the existing SGG methods.

We propose a causal framework to explain the catastrophic forgetting in Class-Incremental Learning (CIL) and then derive a novel distillation method that is orthogonal to the existing anti-forgetting techniques, such as data replay and feature/label distillation. We first 1) place CIL into the framework, 2) answer why the forgetting happens: the causal effect of the old data is lost in new training, and then 3) explain how the existing techniques mitigate it: they bring the causal effect back. Based on the causal framework, we propose to distill the Colliding Effect between the old and the new data, which is fundamentally equivalent to the causal effect of data replay, but without any cost of replay storage. Thanks to the causal effect analysis, we can further capture the Incremental Momentum Effect of the data stream, removing which can help to retain the old effect overwhelmed by the new data effect, and thus alleviate the forgetting of the old class in testing. Extensive experiments on three CIL benchmarks: CIFAR-100, ImageNet-Sub&Full, show that the proposed causal effect distillation can improve various state-of-the-art CIL methods by a large margin (0.72%–9.06%).1

VQA models may tend to rely on language bias as a shortcut and thus fail to sufficiently learn the multi-modal knowledge from both vision and language. Recent debiasing methods proposed to exclude the language prior during inference. However, they fail to disentangle the "good" language context and "bad" language bias from the whole. In this paper, we investigate how to mitigate language bias in VQA. Motivated by causal effects, we proposed a novel counterfactual inference framework, which enables us to capture the language bias as the direct causal effect of questions on answers and reduce the language bias by subtracting the direct language effect from the total causal effect. Experiments demonstrate that our proposed counterfactual inference framework 1) is general to various VQA backbones and fusion strategies, 2) achieves competitive performance on the language-bias sensitive VQA-CP dataset while performs robustly on the balanced VQA v2 dataset without any augmented data. The code is available at https://github.com/yuleiniu/cfvqa.

Real-world visual recognition requires handling the extreme sample imbalance in large-scale long-tailed data. We propose a “divide&conquer” strategy for the challenging LVIS task: divide the whole data into balanced parts and then apply incremental learning to conquer each one. This derives a novel learning paradigm: class-incremental few-shot learning, which is especially effective for the challenge evolving over time: 1) the class imbalance among the old class knowledge review and 2) the few-shot data in new-class learning. We call our approach Learning to Segment the Tail (LST). In particular, we design an instance-level balanced replay scheme, which is a memory-efficient approximation to balance the instance-level samples from the old-class images. We also propose to use a meta-module for new-class learning, where the module parameters are shared across incremental phases, gaining the learning-to-learn knowledge incrementally, from the data-rich head to the data-poor tail. We empirically show that: at the expense of a little sacrifice of head-class forgetting, we can gain a significant 8.3% AP improvement for the tail classes with less than 10 instances, achieving an overall 2.0% AP boost for the whole 1,230 classes.

Today's scene graph generation (SGG) task is still far from practical, mainly due to the severe training bias, e.g., collapsing diverse "human walk on / sit on / lay on beach" into "human on beach". Given such SGG, the down-stream tasks such as VQA can hardly infer better scene structures than merely a bag of objects. However, debiasing in SGG is not trivial because traditional debiasing methods cannot distinguish between the good and bad bias, e.g., good context prior (e.g., "person read book" rather than "eat") and bad long-tailed bias (e.g., "near" dominating "behind / in front of"). In this paper, we present a novel SGG framework based on causal inference but not the conventional likelihood. We first build a causal graph for SGG, and perform traditional biased training with the graph. Then, we propose to draw the counterfactual causality from the trained graph to infer the effect from the bad bias, which should be removed. In particular, we use Total Direct Effect (TDE) as the proposed final predicate score for unbiased SGG. Note that our framework is agnostic to any SGG model and thus can be widely applied in the community who seeks unbiased predictions. By using the proposed Scene Graph Diagnosis toolkit on the SGG benchmark Visual Genome and several prevailing models, we observed significant improvements over the previous state-of-the-art methods.

As the class size grows, maintaining a balanced dataset across many classes is challenging because the data are long-tailed in nature; it is even impossible when the sample-of-interest co-exists with each other in one collectable unit, e.g., multiple visual instances in one image. Therefore, long-tailed classification is the key to deep learning at scale. However, existing methods are mainly based on re-weighting/re-sampling heuristics that lack a fundamental theory. In this paper, we establish a causal inference framework, which not only unravels the whys of previous methods, but also derives a new principled solution. Specifically, our theory shows that the SGD momentum is essentially a confounder in long-tailed classification. On one hand, it has a harmful causal effect that misleads the tail prediction biased towards the head. On the other hand, its induced mediation also benefits the representation learning and head prediction. Our framework elegantly disentangles the paradoxical effects of the momentum, by pursuing the direct causal effect caused by an input sample. In particular, we use causal intervention in training, and counterfactual reasoning in inference, to remove the bad'' while keep thegood''. We achieve new state-of-the-arts on three long-tailed visual recognition benchmarks: Long-tailed CIFAR-10/-100, ImageNet-LT for image classification and LVIS for instance segmentation.

We propose to compose dynamic tree structures that place the objects in an image into a visual context, helping visual reasoning tasks such as scene graph generation and visual Q&A. Our visual context tree model, dubbed VCTree, has two key advantages over existing structured object representations including chains and fully-connected graphs: 1) The efficient and expressive binary tree encodes the inherent parallel/hierarchical relationships among objects, e.g., ``clothes'' and ``pants'' are usually co-occur and belong to ``person''; 2) the dynamic structure varies from image to image and task to task, allowing more content-/task-specific message passing among objects. To construct a VCTree, we design a score function that calculates the task-dependent validity between each object pair, and the tree is the binary version of the maximum spanning tree from the score matrix. Then, visual contexts are encoded by bidirectional TreeLSTM and decoded by task-specific models. We develop a hybrid learning procedure which integrates end-task supervised learning and the tree structure reinforcement learning, where the former's evaluation result serves as a self-critic for the latter's structure exploration. Experimental results on two benchmarks, which require reasoning over contexts: Visual Genome for scene graph generation and VQA2.0 for visual Q&A, show that VCTree outperforms state-of-the-art results while discovering interpretable visual context structures.

We propose Scene Graph Auto-Encoder (SGAE) that incorporates the language inductive bias into the encoder-decoder image captioning framework for more human-like captions. Intuitively, we humans use the inductive bias to compose collocations and contextual inference in discourse. For example, when we see the relation "person on bike'', it is natural to replace "on'' with "ride'' and infer "person riding bike on a road'' even the "road'' is not evident. Therefore, exploiting such bias as a language prior is expected to help the conventional encoder-decoder models less likely to overfit to the dataset bias and focus on reasoning. Specifically, we use the scene graph - a directed graph (G) where an object node is connected by adjective nodes and relationship nodes - to represent the complex structural layout of both image (I) and sentence (S). In the textual domain, we use SGAE to learn a dictionary (D) that helps to reconstruct sentences in the S → G → D → S pipeline, where D encodes the desired language prior; in the vision-language domain, we use the shared D to guide the encoder-decoder in the I → G → D → S pipeline. Thanks to the scene graph representation and shared dictionary, the inductive bias is transferred across domains in principle. We validate the effectiveness of SGAE on the challenging MS-COCO image captioning benchmark, e.g., our SGAE-based single-model achieves a new state-of-the-art 127.8 CIDEr-D on the Karpathy split, and a competitive 125.5 CIDEr-D (c40) on the official server even compared to other ensemble models. Code has been made available at: https://github.com/yangxuntu/SGAE.