ICCV 2023 Papers

The problem of predicting driver attention from the driving perspective is gaining increasing research focus due to its remarkable significance for autonomous driving and assisted driving systems. The driving experience is extremely important for safe driving, a skilled driver is able to effortlessly predict oncoming danger (before it becomes salient) based on the driving experience and quickly pay attention to the corresponding zones. However, the nonobjective driving experience is difficult to model, so a mechanism simulating the driver experience accumulation procedure is absent in existing methods, and the current methods usually follow the technique line of saliency prediction methods to predict driver attention. In this paper, we propose a FeedBack Loop Network (FBLNet), which attempts to model the driving experience accumulation procedure. By over-and-over iterations, FBLNet generates the incremental knowledge that carries rich historically-accumulative and long-term temporal information. The incremental knowledge in our model is like the driving experience of humans. Under the guidance of the incremental knowledge, our model fuses the CNN feature and Transformer feature that are extracted from the input image to predict driver attention. Our model exhibits a solid advantage over existing methods, achieving an outstanding performance improvement on two driver attention benchmark datasets.

View Transformation Module (VTM), where transformations happen between multi-view image features and Bird-Eye-View (BEV) representation, is a crucial step in camera-based BEV perception systems. Currently, the two most prominent VTM paradigms are forward projection and backward projection. Forward projection, represented by Lift-Splat-Shoot, leads to sparsely projected BEV features without post-processing. Backward projection, with BEVFormer being an example, tends to generate false-positive BEV features from incorrect projections due to the lack of utilization on depth. To address the above limitations, we propose a novel forward-backward view transformation module. Our approach compensates for the deficiencies in both existing methods, allowing them to enhance each other to obtain higher quality BEV representations mutually. We instantiate the proposed module with FB-BEV, which achieves a new state-of-the-art result of 62.4% NDS on the nuScenes test set. Code and models are available at https://github.com/NVlabs/FB-BEV

Computer vision datasets frequently contain spurious correlations between task-relevant labels and (easy to learn) latent task-irrelevant attributes (e.g. context). Models trained on such datasets learn "shortcuts" and underperform on bias-conflicting slices of data where the correlation does not hold. In this work, we study the problem of identifying such slices to inform downstream bias mitigation strategies. We propose First Amplify Correlations and Then Slice (FACTS), wherein we first amplify correlations to fit a simple bias-aligned hypothesis via strongly regularized empirical risk minimization. Next, we perform correlation-aware slicing via mixture modeling in bias-aligned feature space to discover underperforming data slices that capture distinct correlations. Despite its simplicity, our method considerably improves over prior work (by as much as 35% precision@10) in correlation bias identification across a range of diverse evaluation settings. Our code is available at https://github.com/yvsriram/FACTS.

Computer vision models have known performance disparities across attributes such as gender and skin tone. This means during tasks such as classification and detection, model performance differs for certain classes based on the demographics of the people in the image. These disparities have been shown to exist, but until now there has not been a unified approach to measure these differences for common use-cases of computer vision models. We present a new benchmark named FACET (FAirness in Computer Vision EvaluaTion), a large, publicly available evaluation set of 32k images for some of the most common vision tasks - image classification, object detection and segmentation. For every image in FACET, we hired expert reviewers to manually annotate person-related attributes such as perceived skin tone and hair type, manually draw bounding boxes and label fine-grained person-related classes such as disk jockey or guitarist. In addition, we use FACET to benchmark state-of-the-art vision models and present a deeper understanding of potential performance disparities and challenges across sensitive demographic attributes. With the exhaustive annotations collected, we probe models using single demographics attributes as well as multiple attributes using an intersectional approach (e.g. hair color and perceived skin tone). Our results show that classification, detection, segmentation, and visual grounding models exhibit performance disparities across demographic attributes and intersections of attributes. These harms suggest that not all people represented in datasets receive fair and equitable treatment in these vision tasks. We hope current and future results using our benchmark will contribute to fairer, more robust vision models. FACET is available publicly at https://facet.metademolab.com

Multi-object tracking (MOT) aims to build moving trajectories for number-agnostic objects. Modern multi-object trackers commonly follow the tracking-by-detection strategy. Therefore, fooling detectors can be an effective solution but it usually requires attacks in multiple successive frames, resulting in low efficiency. Attacking association processes improves efficiency but may require model-specific design, leading to poor generalization. In this paper, we propose a novel False negative and False positive attack (F&F attack) mechanism: it perturbs the input image to erase original detections and to inject deceptive false alarms around original ones while integrating the association attack implicitly. The mechanism can produce effective identity switches against multi-object trackers by only fooling detectors in a few frames. To demonstrate the flexibility of the mechanism, we deploy it to three multi-object trackers (ByteTrack, SORT, and CenterTrack) which are enabled by two representative detectors (YOLOX and CenterNet). Comprehensive experiments on MOT17 and MOT20 datasets show that our method significantly outperforms existing attackers, revealing the vulnerability of the tracking-by-detection paradigm to detection attacks.

Efficient or near-zero-cost proxies were proposed recently to address the demanding computational issues of Neural Architecture Search (NAS) in designing deep neural networks (DNNs), where each candidate architecture network only requires one iteration of backpropagation. The values obtained from proxies are used as predictions of architecture performance for downstream tasks. However, two significant drawbacks hinder the wide adoption of these efficient proxies: 1. they are not adaptive to various NAS search spaces and 2. they are not extensible to multi-modality downstream tasks. To address these two issues, we first propose an Extensible proxy (Eproxy) that utilizes self-supervised, few-shot training to achieve near-zero costs. A key component to our Eproxy's efficiency is the introduction of a barrier layer with randomly initialized frozen convolution parameters, which adds non-linearities to the optimization spaces so that Eproxy can discriminate the performance of architectures at an early stage. We further propose a Discrete Proxy Search (DPS) method to find the optimized training settings for Eproxy with only a handful of benchmarked architectures on the target tasks. Our extensive experiments confirm the effectiveness of both Eproxy and DPS. On the NDS-ImageNet search spaces, Eproxy+DPS achieves a higher average ranking correlation (Spearman r = 0.73) than the previous efficient proxy (Spearman r = 0.56). On the NAS-Bench-Trans-Micro search spaces with seven tasks, Eproxy+DPS delivers comparable performance with the early stopping method (146x faster). For the end-to-end task such as DARTS-ImageNet-1k, our method delivers better results than NAS performed on CIFAR-10 while only requiring one GPU hour with a single batch of CIFAR-10 images.

Plain text has become a prevalent interface for text-to-image synthesis. However, its limited customization options hinder users from accurately describing desired outputs. For example, plain text makes it hard to specify continuous quantities, such as the precise RGB color value or importance of each word. Furthermore, creating detailed text prompts for complex scenes is tedious for humans to write and challenging for text encoders to interpret. To address these challenges, we propose using a rich-text editor supporting formats such as font style, size, color, and footnote. We extract each word's attributes from rich text to enable local style control, explicit token reweighting, precise color rendering, and detailed region synthesis. We achieve these capabilities through a region-based diffusion process. We first obtain each word's region based on attention maps of a diffusion process using plain text. For each region, we enforce its text attributes by creating region-specific detailed prompts and applying region-specific guidance, and maintain its fidelity against plain-text generation through region-based injections. We present various examples of image generation from rich text and demonstrate that our method outperforms strong baselines with quantitative evaluations.

Previous raw image-based low-light image enhancement methods predominantly relied on feed-forward neural networks to learn deterministic mappings from low-light to normally-exposed images. However, they failed to capture critical distribution information, leading to visually undesirable results. This work addresses the issue by seamlessly integrating a diffusion model with a physics-based exposure model. Different from a vanilla diffusion model that has to perform Gaussian denoising, with the injected physics-based exposure model, our restoration process can directly start from a noisy image instead of pure noise. As such, our method obtains significantly improved performance and reduced inference time compared with vanilla diffusion models. To make full use of the advantages of different intermediate steps, we further propose an adaptive residual layer that effectively screens out the side-effect in the iterative refinement when the intermediate results have been already well-exposed. The proposed framework can work with both real-paired datasets, SOTA noise models, and different backbone networks. We evaluate the proposed method on various public benchmarks, achieving promising results with consistent improvements using different exposure models and backbones. Besides, the proposed method achieves better generalization capacity for unseen amplifying ratios and better performance than a larger feedforward neural model when few parameters are adopted. The code is released at https://github.com/wyf0912/ExposureDiffusion.

It is very time consuming to create datasets for training computer vision models. An emerging alternative is to use synthetic data, but if the synthetic data is not similar enough to the real data, the performance is typically below that of training with real data. Thus using synthetic data still requires a large amount of time, money, and skill as one needs to author the data carefully. In this paper, we seek to understand which aspects of this authoring process are most critical. We present an analysis of which factors of variation between simulated and real data are most important. We capture images of YCB objects to create a novel YCB-Real dataset. We then create a novel synthetic "digital twin" dataset, YCB-Synthetic, which matches the YCB-Real dataset and includes variety of artifacts added to the synthetic data. We study the affects of these artifacts on our dataset and two existing published datasets on two different computer vision tasks: object detection and instance segmentation. We provide an analysis of the cost-benefit trade-offs between artist time for fixing artifacts and trained model accuracy. We plan to release this dataset (images and 3D assets) so they can be further used by the community.

Visual Prompting (VP) is an emerging and powerful technique that allows sample-efficient adaptation to downstream tasks by engineering a well-trained frozen source model. In this work, we explore the benefits of VP in constructing compelling neural network classifiers with differential privacy (DP). We explore and integrate VP into canonical DP training methods and demonstrate its simplicity and efficiency. In particular, we discover that VP in tandem with PATE, a state-of-the-art DP training method that leverages the knowledge transfer from an ensemble of teachers, achieves the state-of-the-art privacy-utility trade-off with minimum expenditure of privacy budget. Moreover, we conduct additional experiments on cross-domain image classification with a sufficient domain gap to further unveil the advantage of VP in DP. Lastly, we also conduct extensive ablation studies to validate the effectiveness and contribution of VP under DP consideration.

The rapid increase in user-generated-content (UGC) videos calls for the development of effective video quality assessment (VQA) algorithms. However, the objective of the UGC-VQA problem is still ambiguous and can be viewed from two perspectives: the technical perspective, measuring the perception of distortions; and the aesthetic perspective, which relates to preference and recommendation on contents. To understand how these two perspectives affect overall subjective opinions in UGC-VQA, we conduct a large-scale subjective study to collect human quality opinions on overall quality of videos as well as perceptions from aesthetic and technical perspectives. The collected Disentangled Video Quality Database (DIVIDE-3k) confirms that human quality opinions on UGC videos are universally and inevitably affected by both aesthetic and technical perspectives. In light of this, we propose the Disentangled Objective Video Quality Evaluator (DOVER) to learn the quality of UGC videos based on the two perspectives. The DOVER proves state-of-the-art performance in UGC-VQA under very high efficiency. With perspective opinions in DIVIDE-3k, we further propose DOVER++, the first approach to provide reliable clear-cut quality evaluations from a single aesthetic or technical perspective.

Open-world instance segmentation is a rising task, which aims to segment all objects in the image by learning from a limited number of base-category objects. This task is challenging, as the number of unseen categories could be hundreds of times larger than that of seen categories. Recently, the DETR-like models have been extensively studied in the closed world while stay unexplored in the open world. In this paper, we utilize the Transformer for open-world instance segmentation and present SWORD. Firstly, we introduce to attach the stop-gradient operation before classification head and further add IoU heads for discovering novel objects. We demonstrate that a simple stop-gradient operation not only prevents the novel objects from being suppressed as background, but also allows the network to enjoy the merit of heuristic label assignment. Secondly, we propose a novel contrastive learning framework to enlarge the representations between objects and background. Specifically, we maintain a universal object queue to obtain the object center, and dynamically select positive and negative samples from the object queries for contrastive learning. While the previous works only focus on pursuing average recall and neglect average precision, we show the prominence of SWORD by giving consideration to both criteria. Our models achieve state-of-the-art performance in various open-world cross-category and cross-dataset generalizations. Particularly, in VOC to non-VOC setup, our method sets new state-of-the-art results of 40.0% on ARb100 and 34.9% on ARm100. For COCO to UVO generalization, SWORD significantly outperforms the previous best open-world model by 5.9% on APm and 8.1% on ARm100, respectively.

Video deblurring aims to restore the latent video frames from their blurred counterparts. Despite the remarkable progress, most promising video deblurring methods only investigate the temporal priors in the spatial domain and rarely explore their its potential in the frequency domain. In this paper, we revisit the blurred sequence in the Fourier space and figure out some intrinsic frequency-temporal priors that imply the temporal blur degradation can be accessibly decoupled in the potential frequency domain. Based on these priors, we propose a novel Fourier-based frequency-temporal video deblurring solution, where the core design accommodates the temporal spectrum to a popular video deblurring pipeline of feature extraction, alignment, aggregation, and optimization. Specifically, we design a Spectrum Prior-guided Alignment module by leveraging enlarged blur information in the potential spectrum to mitigate the blur effects on the alignment. Then, Temporal Energy prior-driven Aggregation is implemented to replenish the original local features by estimating the temporal spectrum energy as the global sharpness guidance. In addition, the customized frequency loss is devised to optimize the proposed method for decent spectral distribution. Extensive experiments demonstrate that our model performs favorably against other state-of-the-art methods, thus confirming the effectiveness of frequency-temporal prior modeling.

Paired video and language data is naturally temporal concurrency, which requires the modeling of the temporal dynamics within each modality and the temporal alignment across modalities simultaneously. However, most existing video-language representation learning methods only focus on discrete semantic alignment that encourages aligned semantics to be close in the latent space, or temporal context dependency that captures short-range coherence, failing in building the temporal concurrency. In this paper, we propose to learn video-language representations by modeling video-language pairs as Temporal Concurrent Processes (TCP) via a process-wised distance metric learning framework. Specifically, we employ the soft Dynamic Time Warping (DTW) to measure the distance between two processes across modalities and then optimize the DTW costs. Meanwhile, we further introduce a regularization term that enforces the embeddings of each modality approximating a stochastic process to guarantee the inherent dynamics. Experimental results on three benchmarks demonstrate that TCP stands as a state-of-the-art method for various video-language understanding tasks, including paragraph-to-video retrieval, video moment retrieval, and video question-answering. Code is available at https://github.com/hengRUC/TCP.

Recently, the DETR framework has emerged as the dominant approach for human--object interaction (HOI) research. In particular, two-stage transformer-based HOI detectors are amongst the most performant and training-efficient approaches. However, these often condition HOI classification on object features that lack fine-grained contextual information, eschewing pose and orientation information in favour of visual cues about object identity and box extremities. This naturally hinders the recognition of complex or ambiguous interactions. In this work, we study these issues through visualisations and carefully designed experiments. Accordingly, we investigate how best to re-introduce image features via cross-attention. With an improved query design, extensive exploration of keys and values, and box pair positional embeddings as spatial guidance, our model with enhanced predicate visual context (PViC) outperforms state-of-the-art methods on the HICO-DET and V-COCO benchmarks, while maintaining low training cost.

In digital photography, autofocus is a key feature that aids high-quality image capture, and modern approaches use the phase patterns arising from dual-pixel sensors as important focus cues. However, dual-pixel data is prone to multiple error sources in its image capturing process, including lens shading or distortions due to the inherent optical characteristics of the lens. We observe that, while these degradations are hard to model using prior knowledge, they are correlated with the spatial position of the pixels within the image sensor area, and we propose a learning-based autofocus model with positional encodings (PE) to capture these patterns. Specifically, we introduce RoI-PE, which encodes the spatial position of our focusing region-of-interest (RoI) on the imaging plane. Learning with RoI-PE allows the model to be more robust to spatially-correlated degradations. In addition, we also propose to encode the current focal position of lens as lens-PE, which allows us to significantly reduce the computational complexity of the autofocus model. Experimental results clearly demonstrate the effectiveness of using the proposed position encodings for automatic focusing based on dual-pixel data.

Semantic segmentation is a crucial task in computer vision that involves segmenting images into semantically meaningful regions at the pixel level. However, existing approaches often rely on expensive human annotations as supervision for model training, limiting their scalability to large, unlabeled datasets. To address this challenge, we present ZeroSeg, a novel method that leverages the existing pretrained vision-language (VL) model (e.g. CLIP vision encoder) to train open-vocabulary zero-shot semantic segmentation models. Although acquired extensive knowledge of visual concepts, it is non-trivial to exploit knowledge from these VL models to the task of semantic segmentation, as they are usually trained at an image level. ZeroSeg overcomes this by distilling the visual concepts learned by VL models into a set of segment tokens, each summarizing a localized region of the target image. We evaluate ZeroSeg on multiple popular segmentation benchmarks, including PASCAL VOC 2012, PASCAL Context, and COCO, in a zero-shot manner Our approach achieves state-of-the-art performance when compared to other zero-shot segmentation methods under the same training data, while also performing competitively compared to strongly supervised methods. Finally, we also demonstrated the effectiveness of ZeroSeg on open-vocabulary segmentation, through both human studies and qualitative visualizations. The code is publicly available at https://github.com/facebookresearch/ZeroSeg

In this paper, we propose a long-sequence modeling framework, named StreamPETR, for multi-view 3D object detection. Built upon the sparse query design in the PETR series, we systematically develop an object-centric temporal mechanism. The model is performed in an online manner and the long-term historical information is propagated through object queries frame by frame. Besides, we introduce a motion-aware layer normalization to model the movement of the objects. StreamPETR achieves significant performance improvements only with negligible computation cost, compared to the single-frame baseline. On the standard nuScenes benchmark, it is the first online multi-view method that achieves comparable performance (67.6% NDS & 65.3% AMOTA) with lidar-based methods. The lightweight version realizes 45.0% mAP and 31.7 FPS, outperforming the state-of-the-art method (SOLOFusion) by 2.3% mAP and 1.8x faster FPS. Code has been available at https://github.com/exiawsh/StreamPETR.git.

Transfer learning has become crucial in computer vision tasks due to the vast availability of pre-trained deep learning models. However, selecting the optimal pre-trained model from a diverse pool for a specific downstream task remains a challenge. Existing methods for measuring the transferability of pre-trained models rely on statistical correlations between encoded static features and task labels, but they overlook the impact of underlying representation dynamics during fine-tuning, leading to unreliable results, especially for self-supervised models. In this paper, we present an insightful physics-inspired approach named PED to address these challenges. We reframe the challenge of model selection through the lens of potential energy and directly model the interaction forces that influence fine-tuning dynamics. By capturing the motion of dynamic representations to decline the potential energy within a force-driven physical model, we can acquire an enhanced and more stable observation for estimating transferability. The experimental results on 10 downstream tasks and 12 self-supervised models demonstrate that our approach can seamlessly integrate into existing ranking techniques and enhance their performances, revealing its effectiveness for the model selection task and its potential for understanding the mechanism in transfer learning. Code is available at https://github.com/lixiaotong97/PED.