CVPR 2023 论文列表

Recent advances in 3D object detection (3DOD) have obtained remarkably strong results for LiDAR-based models. In contrast, surround-view 3DOD models based on multiple camera images underperform due to the necessary view transformation of features from perspective view (PV) to a 3D world representation which is ambiguous due to missing depth information. This paper introduces X3KD, a comprehensive knowledge distillation framework across different modalities, tasks, and stages for multi-camera 3DOD. Specifically, we propose cross-task distillation from an instance segmentation teacher (X-IS) in the PV feature extraction stage providing supervision without ambiguous error backpropagation through the view transformation. After the transformation, we apply cross-modal feature distillation (X-FD) and adversarial training (X-AT) to improve the 3D world representation of multi-camera features through the information contained in a LiDAR-based 3DOD teacher. Finally, we also employ this teacher for cross-modal output distillation (X-OD), providing dense supervision at the prediction stage. We perform extensive ablations of knowledge distillation at different stages of multi-camera 3DOD. Our final X3KD model outperforms previous state-of-the-art approaches on the nuScenes and Waymo datasets and generalizes to RADAR-based 3DOD. Qualitative results video at https://youtu.be/1do9DPFmr38.

Weak-shot Object Detection methods exploit a fully-annotated source dataset to facilitate the detection performance on the target dataset which only contains image-level labels for novel categories. To bridge the gap between these two datasets, we aim to transfer the object knowledge between the source (S) and target (T) datasets in a bi-directional manner. We propose a novel Knowledge Transfer (KT) loss which simultaneously distills the knowledge of objectness and class entropy from a proposal generator trained on the S dataset to optimize a multiple instance learning module on the T dataset. By jointly optimizing the classification loss and the proposed KT loss, the multiple instance learning module effectively learns to classify object proposals into novel categories in the T dataset with the transferred knowledge from base categories in the S dataset. Noticing the predicted boxes on the T dataset can be regarded as an extension for the original annotations on the S dataset to refine the proposal generator in return, we further propose a novel Consistency Filtering (CF) method to reliably remove inaccurate pseudo labels by evaluating the stability of the multiple instance learning module upon noise injections. Via mutually transferring knowledge between the S and T datasets in an iterative manner, the detection performance on the target dataset is significantly improved. Extensive experiments on public benchmarks validate that the proposed method performs favourably against the state-of-the-art methods without increasing the model parameters or inference computational complexity.

In the field of 3D object detection for autonomous driving, the sensor portfolio including multi-modality and single-modality is diverse and complex. Since the multi-modal methods have system complexity while the accuracy of single-modal ones is relatively low, how to make a tradeoff between them is difficult. In this work, we propose a universal cross-modality knowledge distillation framework (UniDistill) to improve the performance of single-modality detectors. Specifically, during training, UniDistill projects the features of both the teacher and the student detector into Bird's-Eye-View (BEV), which is a friendly representation for different modalities. Then, three distillation losses are calculated to sparsely align the foreground features, helping the student learn from the teacher without introducing additional cost during inference. Taking advantage of the similar detection paradigm of different detectors in BEV, UniDistill easily supports LiDAR-to-camera, camera-to-LiDAR, fusion-to-LiDAR and fusion-to-camera distillation paths. Furthermore, the three distillation losses can filter the effect of misaligned background information and balance between objects of different sizes, improving the distillation effectiveness. Extensive experiments on nuScenes demonstrate that UniDistill effectively improves the mAP and NDS of student detectors by 2.0% 3.2%.

How to estimate the uncertainty of a given model is a crucial problem. Current calibration techniques treat different classes equally and thus implicitly assume that the distribution of training data is balanced, but ignore the fact that real-world data often follows a long-tailed distribution. In this paper, we explore the problem of calibrating the model trained from a long-tailed distribution. Due to the difference between the imbalanced training distribution and balanced test distribution, existing calibration methods such as temperature scaling can not generalize well to this problem. Specific calibration methods for domain adaptation are also not applicable because they rely on unlabeled target domain instances which are not available. Models trained from a long-tailed distribution tend to be more overconfident to head classes. To this end, we propose a novel knowledge-transferring-based calibration method by estimating the importance weights for samples of tail classes to realize long-tailed calibration. Our method models the distribution of each class as a Gaussian distribution and views the source statistics of head classes as a prior to calibrate the target distributions of tail classes. We adaptively transfer knowledge from head classes to get the target probability density of tail classes. The importance weight is estimated by the ratio of the target probability density over the source probability density. Extensive experiments on CIFAR-10-LT, MNIST-LT, CIFAR-100-LT, and ImageNet-LT datasets demonstrate the effectiveness of our method.

Video captioning aims to describe the content of videos using natural language. Although significant progress has been made, there is still much room to improve the performance for real-world applications, mainly due to the long-tail and open set issues of words. In this paper, we propose a text with knowledge graph augmented transformer (TextKG) for video captioning. Notably, TextKG is a two-stream transformer, formed by the external stream and internal stream. The external stream is designed to absorb external knowledge, which models the interactions between the external knowledge, e.g., pre-built knowledge graph, and the built-in information of videos, e.g., the salient object regions, speech transcripts, and video captions, to mitigate the open set of words challenge. Meanwhile, the internal stream is designed to exploit the multi-modality information in original videos (e.g., the appearance of video frames, speech transcripts, and video captions) to deal with the long-tail issue. In addition, the cross attention mechanism is also used in both streams to share information. In this way, the two streams can help each other for more accurate results. Extensive experiments conducted on four challenging video captioning datasets, i.e., YouCookII, ActivityNet Captions, MSR-VTT, and MSVD, demonstrate that the proposed method performs favorably against the state-of-the-art methods. Specifically, the proposed TextKG method outperforms the best published results by improving 18.7% absolute CIDEr scores on the YouCookII dataset.

Visual place recognition (VPR) is usually considered as a specific image retrieval problem. Limited by existing training frameworks, most deep learning-based works cannot extract sufficiently stable global features from RGB images and rely on a time-consuming re-ranking step to exploit spatial structural information for better performance. In this paper, we propose StructVPR, a novel training architecture for VPR, to enhance structural knowledge in RGB global features and thus improve feature stability in a constantly changing environment. Specifically, StructVPR uses segmentation images as a more definitive source of structural knowledge input into a CNN network and applies knowledge distillation to avoid online segmentation and inference of seg-branch in testing. Considering that not all samples contain high-quality and helpful knowledge, and some even hurt the performance of distillation, we partition samples and weigh each sample's distillation loss to enhance the expected knowledge precisely. Finally, StructVPR achieves impressive performance on several benchmarks using only global retrieval and even outperforms many two-stage approaches by a large margin. After adding additional re-ranking, ours achieves state-of-the-art performance while maintaining a low computational cost.

Existing methods mainly handle single weather types. However, the connections of different weather conditions at deep representation level are usually ignored. These connections, if used properly, can generate complementary representations for each other to make up insufficient training data, obtaining positive performance gains and better generalization. In this paper, we focus on the very correlated rain and snow to explore their connections at deep representation level. Because sub-optimal connections may cause negative effect, another issue is that if rain and snow are handled in a multi-task learning way, how to find an optimal connection strategy to simultaneously improve deraining and desnowing performance. To build desired connection, we propose a smart knowledge assignment strategy, called SmartAssign, to optimally assign the knowledge learned from both tasks to a specific one. In order to further enhance the accuracy of knowledge assignment, we propose a novel knowledge contrast mechanism, so that the knowledge assigned to different tasks preserves better uniqueness. The inherited inductive biases usually limit the modelling ability of CNNs, we introduce a novel transformer block to constitute the backbone of our network to effectively combine long-range context dependency and local image details. Extensive experiments on seven benchmark datasets verify that proposed SmartAssign explores effective connection between rain and snow, and improves the performances of both deraining and desnowing apparently. The implementation code will be available at https://gitee.com/mindspore/models/tree/master/research/cv/SmartAssign.

Modern Generative Adversarial Networks (GANs) generate realistic images remarkably well. Previous work has demonstrated the feasibility of "GAN-classifiers" that are distinct from the co-trained discriminator, and operate on images generated from a frozen GAN. That such classifiers work at all affirms the existence of "knowledge gaps" (out-of-distribution artifacts across samples) present in GAN training. We iteratively train GAN-classifiers and train GANs that "fool" the classifiers (in an attempt to fill the knowledge gaps), and examine the effect on GAN training dynamics, output quality, and GAN-classifier generalization. We investigate two settings, a small DCGAN architecture trained on low dimensional images (MNIST), and StyleGAN2, a SOTA GAN architecture trained on high dimensional images (FFHQ). We find that the DCGAN is unable to effectively fool a held-out GAN-classifier without compromising the output quality. However, StyleGAN2 can fool held-out classifiers with no change in output quality, and this effect persists over multiple rounds of GAN/classifier training which appears to reveal an ordering over optima in the generator parameter space. Finally, we study different classifier architectures and show that the architecture of the GAN-classifier has a strong influence on the set of its learned artifacts.

This paper demonstrates an approach for learning highly semantic image representations without relying on hand-crafted data-augmentations. We introduce the Image-based Joint-Embedding Predictive Architecture (I-JEPA), a non-generative approach for self-supervised learning from images. The idea behind I-JEPA is simple: from a single context block, predict the representations of various target blocks in the same image. A core design choice to guide I-JEPA towards producing semantic representations is the masking strategy; specifically, it is crucial to (a) sample target blocks with sufficiently large scale (semantic), and to (b) use a sufficiently informative (spatially distributed) context block. Empirically, when combined with Vision Transformers, we find I-JEPA to be highly scalable. For instance, we train a ViT-Huge/14 on ImageNet using 16 A100 GPUs in under 72 hours to achieve strong downstream performance across a wide range of tasks, from linear classification to object counting and depth prediction.

Despite the prominent success of general object detection, the performance and efficiency of Small Object Detection (SOD) are still unsatisfactory. Unlike existing works that struggle to balance the trade-off between inference speed and SOD performance, in this paper, we propose a novel Scale-aware Knowledge Distillation (ScaleKD), which transfers knowledge of a complex teacher model to a compact student model. We design two novel modules to boost the quality of knowledge transfer in distillation for SOD: 1) a scale-decoupled feature distillation module that disentangled teacher's feature representation into multi-scale embedding that enables explicit feature mimicking of the student model on small objects. 2) a cross-scale assistant to refine the noisy and uninformative bounding boxes prediction student models, which can mislead the student model and impair the efficacy of knowledge distillation. A multi-scale cross-attention layer is established to capture the multi-scale semantic information to improve the student model. We conduct experiments on COCO and VisDrone datasets with diverse types of models, i.e., two-stage and one-stage detectors, to evaluate our proposed method. Our ScaleKD achieves superior performance on general detection performance and obtains spectacular improvement regarding the SOD performance.

With the continual expansion of face datasets, feature-based distillation prevails for large-scale face recognition. In this work, we attempt to remove identity supervision in student training, to spare the GPU memory from saving massive class centers. However, this naive removal leads to inferior distillation result. We carefully inspect the performance degradation from the perspective of intrinsic dimension, and argue that the gap in intrinsic dimension, namely the intrinsic gap, is intimately connected to the infamous capacity gap problem. By constraining the teacher's search space with reverse distillation, we narrow the intrinsic gap and unleash the potential of feature-only distillation. Remarkably, the proposed reverse distillation creates universally student-friendly teacher that demonstrates outstanding student improvement. We further enhance its effectiveness by designing a student proxy to better bridge the intrinsic gap. As a result, the proposed method surpasses state-of-the-art distillation techniques with identity supervision on various face recognition benchmarks, and the improvements are consistent across different teacher-student pairs.

In this paper, we propose an end-to-end Retrieval-Augmented Visual Language Model (REVEAL) that learns to encode world knowledge into a large-scale memory, and to retrieve from it to answer knowledge-intensive queries. REVEAL consists of four key components: the memory, the encoder, the retriever and the generator. The large-scale memory encodes various sources of multimodal world knowledge (e.g. image-text pairs, question answering pairs, knowledge graph triplets, etc.) via a unified encoder. The retriever finds the most relevant knowledge entries in the memory, and the generator fuses the retrieved knowledge with the input query to produce the output. A key novelty in our approach is that the memory, encoder, retriever and generator are all pre-trained end-to-end on a massive amount of data. Furthermore, our approach can use a diverse set of multimodal knowledge sources, which is shown to result in significant gains. We show that REVEAL achieves state-of-the-art results on visual question answering and image captioning.

Knowledge-based visual question answering (VQA) requires external knowledge beyond the image to answer the question. Early studies retrieve required knowledge from explicit knowledge bases (KBs), which often introduces irrelevant information to the question, hence restricting the performance of their models. Recent works have sought to use a large language model (i.e., GPT-3) as an implicit knowledge engine to acquire the necessary knowledge for answering. Despite the encouraging results achieved by these methods, we argue that they have not fully activated the capacity of GPT-3 as the provided input information is insufficient. In this paper, we present Prophet---a conceptually simple framework designed to prompt GPT-3 with answer heuristics for knowledge-based VQA. Specifically, we first train a vanilla VQA model on a specific knowledge-based VQA dataset without external knowledge. After that, we extract two types of complementary answer heuristics from the model: answer candidates and answer-aware examples. Finally, the two types of answer heuristics are encoded into the prompts to enable GPT-3 to better comprehend the task thus enhancing its capacity. Prophet significantly outperforms all existing state-of-the-art methods on two challenging knowledge-based VQA datasets, OK-VQA and A-OKVQA, delivering 61.1% and 55.7% accuracies on their testing sets, respectively.

Mean ensemble (i.e. averaging predictions from multiple models) is a commonly-used technique in machine learning that improves the performance of each individual model. We formalize it as feature alignment for ensemble in open-set face recognition and generalize it into Bayesian Ensemble Averaging (BEA) through the lens of probabilistic modeling. This generalization brings up two practical benefits that existing methods could not provide: (1) the uncertainty of a face image can be evaluated and further decomposed into aleatoric uncertainty and epistemic uncertainty, the latter of which can be used as a measure for out-of-distribution detection of faceness; (2) a BEA statistic provably reflects the aleatoric uncertainty of a face image, acting as a measure for face image quality to improve recognition performance. To inherit the uncertainty estimation capability from BEA without the loss of inference efficiency, we propose BEA-KD, a student model to distill knowledge from BEA. BEA-KD mimics the overall behavior of ensemble members and consistently outperforms SOTA knowledge distillation methods on various challenging benchmarks.

The quality of scene graphs generated by the state-of-the-art (SOTA) models is compromised due to the long-tail nature of the relationships and their parent object pairs. Training of the scene graphs is dominated by the majority relationships of the majority pairs and, therefore, the object-conditional distributions of relationship in the minority pairs are not preserved after the training is converged. Consequently, the biased model performs well on more frequent relationships in the marginal distribution of relationships such as 'on' and 'wearing', and performs poorly on the less frequent relationships such as 'eating' or 'hanging from'. In this work, we propose virtual evidence incorporated within-triplet Bayesian Network (BN) to preserve the object-conditional distribution of the relationship label and to eradicate the bias created by the marginal probability of the relationships. The insufficient number of relationships in the minority classes poses a significant problem in learning the within-triplet Bayesian network. We address this insufficiency by embedding-based augmentation of triplets where we borrow samples of the minority triplet classes from its neighboring triplets in the semantic space. We perform experiments on two different datasets and achieve a significant improvement in the mean recall of the relationships. We also achieve a better balance between recall and mean recall performance compared to the SOTA de-biasing techniques of scene graph models.

In this paper, we study a novel task that enables partial knowledge transfer from pre-trained models, which we term as Partial Network Cloning (PNC). Unlike prior methods that update all or at least part of the parameters in the target network throughout the knowledge transfer process, PNC conducts partial parametric "cloning" from a source network and then injects the cloned module to the target, without modifying its parameters. Thanks to the transferred module, the target network is expected to gain additional functionality, such as inference on new classes; whenever needed, the cloned module can be readily removed from the target, with its original parameters and competence kept intact. Specifically, we introduce an innovative learning scheme that allows us to identify simultaneously the component to be cloned from the source and the position to be inserted within the target network, so as to ensure the optimal performance. Experimental results on several datasets demonstrate that, our method yields a significant improvement of 5% in accuracy and 50% in locality when compared with parameter-tuning based methods.

Self-supervised learning (SSL) has made remarkable progress in visual representation learning. Some studies combine SSL with knowledge distillation (SSL-KD) to boost the representation learning performance of small models. In this study, we propose a Multi-mode Online Knowledge Distillation method (MOKD) to boost self-supervised visual representation learning. Different from existing SSL-KD methods that transfer knowledge from a static pre-trained teacher to a student, in MOKD, two different models learn collaboratively in a self-supervised manner. Specifically, MOKD consists of two distillation modes: self-distillation and cross-distillation modes. Among them, self-distillation performs self-supervised learning for each model independently, while cross-distillation realizes knowledge interaction between different models. In cross-distillation, a cross-attention feature search strategy is proposed to enhance the semantic feature alignment between different models. As a result, the two models can absorb knowledge from each other to boost their representation learning performance. Extensive experimental results on different backbones and datasets demonstrate that two heterogeneous models can benefit from MOKD and outperform their independently trained baseline. In addition, MOKD also outperforms existing SSL-KD methods for both the student and teacher models.

The archaeological dating of bronze dings has played a critical role in the study of ancient Chinese history. Current archaeology depends on trained experts to carry out bronze dating, which is time-consuming and labor-intensive. For such dating, in this study, we propose a learning-based approach to integrate advanced deep learning techniques and archaeological knowledge. To achieve this, we first collect a large-scale image dataset of bronze dings, which contains richer attribute information than other existing fine-grained datasets. Second, we introduce a multihead classifier and a knowledge-guided relation graph to mine the relationship between attributes and the ding era. Third, we conduct comparison experiments with various existing methods, the results of which show that our dating method achieves a state-of-the-art performance. We hope that our data and applied networks will enrich fine-grained classification research relevant to other interdisciplinary areas of expertise. The dataset and source code used are included in our supplementary materials, and will be open after submission owing to the anonymity policy. Source codes and data are available at: https://github.com/zhourixin/bronze-Ding.

This paper studies the potential of distilling knowledge from pre-trained models, especially Masked Autoencoders. Our approach is simple: in addition to optimizing the pixel reconstruction loss on masked inputs, we minimize the distance between the intermediate feature map of the teacher model and that of the student model. This design leads to a computationally efficient knowledge distillation framework, given 1) only a small visible subset of patches is used, and 2) the (cumbersome) teacher model only needs to be partially executed, i.e., forward propagate inputs through the first few layers, for obtaining intermediate feature maps. Compared to directly distilling fine-tuned models, distilling pre-trained models substantially improves downstream performance. For example, by distilling the knowledge from an MAE pre-trained ViT-L into a ViT-B, our method achieves 84.0% ImageNet top-1 accuracy, outperforming the baseline of directly distilling a fine-tuned ViT-L by 1.2%. More intriguingly, our method can robustly distill knowledge from teacher models even with extremely high masking ratios: e.g., with 95% masking ratio where merely TEN patches are visible during distillation, our ViT-B competitively attains a top-1 ImageNet accuracy of 83.6%; surprisingly, it can still secure 82.4% top-1 ImageNet accuracy by aggressively training with just FOUR visible patches (98% masking ratio). The code will be made publicly available.