CVPR 2023 Papers

Visual Question Answering models have been shown to suffer from language biases, where the model learns a correlation between the question and the answer, ignoring the image. While early works attempted to use question-only models or data augmentations to reduce this bias, we propose an adaptive margin loss approach having two components. The first component considers the frequency of answers within a question type in the training data, which addresses the concern of the class-imbalance causing the language biases. However, it does not take into account the answering difficulty of the samples, which impacts their learning. We address this through the second component, where instance-specific margins are learnt, allowing the model to distinguish between samples of varying complexity. We introduce a bias-injecting component to our model, and compute the instance-specific margins from the confidence of this component. We combine these with the estimated margins to consider both answer-frequency and task-complexity in the training loss. We show that, while the margin loss is effective for out-of-distribution (ood) data, the bias-injecting component is essential for generalising to in-distribution (id) data. Our proposed approach, Robust Margin Loss for Visual Question Answering (RMLVQA) improves upon the existing state-of-the-art results when compared to augmentation-free methods on benchmark VQA datasets suffering from language biases, while maintaining competitive performance on id data, making our method the most robust one among all comparable methods.

Both masked image modeling (MIM) and natural language supervision have facilitated the progress of transferable visual pre-training. In this work, we seek the synergy between two paradigms and study the emerging properties when MIM meets natural language supervision. To this end, we present a novel masked visual Reconstruction In Language semantic Space (RILS) pre-training framework, in which sentence representations, encoded by the text encoder, serve as prototypes to transform the vision-only signals into patch-sentence probabilities as semantically meaningful MIM reconstruction targets. The vision models can therefore capture useful components with structured information by predicting proper semantic of masked tokens. Better visual representations could, in turn, improve the text encoder via the image-text alignment objective, which is essential for the effective MIM target transformation. Extensive experimental results demonstrate that our method not only enjoys the best of previous MIM and CLIP but also achieves further improvements on various tasks due to their mutual benefits. RILS exhibits advanced transferability on downstream classification, detection, and segmentation, especially for low-shot regimes. Code is available at https://github.com/hustvl/RILS.

This paper studies how to keep a vision backbone effective while removing token mixers in its basic building blocks. Token mixers, as self-attention for vision transformers (ViTs), are intended to perform information communication between different spatial tokens but suffer from considerable computational cost and latency. However, directly removing them will lead to an incomplete model structure prior, and thus brings a significant accuracy drop. To this end, we first develop an RepIdentityFormer base on the re-parameterizing idea, to study the token mixer free model architecture. And we then explore the improved learning paradigm to break the limitation of simple token mixer free backbone, and summarize the empirical practice into 5 guidelines. Equipped with the proposed optimization strategy, we are able to build an extremely simple vision backbone with encouraging performance, while enjoying the high efficiency during inference. Extensive experiments and ablative analysis also demonstrate that the inductive bias of network architecture, can be incorporated into simple network structure with appropriate optimization strategy. We hope this work can serve as a starting point for the exploration of optimization-driven efficient network design.

Existing dehazing approaches struggle to process real-world hazy images owing to the lack of paired real data and robust priors. In this work, we present a new paradigm for real image dehazing from the perspectives of synthesizing more realistic hazy data and introducing more robust priors into the network. Specifically, (1) instead of adopting the de facto physical scattering model, we rethink the degradation of real hazy images and propose a phenomenological pipeline considering diverse degradation types. (2) We propose a Real Image Dehazing network via high-quality Codebook Priors (RIDCP). Firstly, a VQGAN is pre-trained on a large-scale high-quality dataset to obtain the discrete codebook, encapsulating high-quality priors (HQPs). After replacing the negative effects brought by haze with HQPs, the decoder equipped with a novel normalized feature alignment module can effectively utilize high-quality features and produce clean results. However, although our degradation pipeline drastically mitigates the domain gap between synthetic and real data, it is still intractable to avoid it, which challenges HQPs matching in the wild. Thus, we re-calculate the distance when matching the features to the HQPs by a controllable matching operation, which facilitates finding better counterparts. We provide a recommendation to control the matching based on an explainable solution. Users can also flexibly adjust the enhancement degree as per their preference. Extensive experiments verify the effectiveness of our data synthesis pipeline and the superior performance of RIDCP in real image dehazing. Code and data will be released.

This paper presents a learning-based method for multi-view depth estimation from posed images. Our core idea is a "learning-to-optimize" paradigm that iteratively indexes a plane-sweeping cost volume and regresses the depth map via a convolutional Gated Recurrent Unit (GRU). Since the cost volume plays a paramount role in encoding the multi-view geometry, we aim to improve its construction both at pixel- and frame- levels. At the pixel level, we propose to break the symmetry of the Siamese network (which is typically used in MVS to extract image features) by introducing a transformer block to the reference image (but not to the source images). Such an asymmetric volume allows the network to extract global features from the reference image to predict its depth map. Given potential inaccuracies in the poses between reference and source images, we propose to incorporate a residual pose network to correct the relative poses. This essentially rectifies the cost volume at the frame level. We conduct extensive experiments on real-world MVS datasets and show that our method achieves state-of-the-art performance in terms of both within-dataset evaluation and cross-dataset generalization.

The field of text-to-image generation has made remarkable strides in creating high-fidelity and photorealistic images. As this technology gains popularity, there is a growing concern about its potential security risks. However, there has been limited exploration into the robustness of these models from an adversarial perspective. Existing research has primarily focused on untargeted settings, and lacks holistic consideration for reliability (attack success rate) and stealthiness (imperceptibility). In this paper, we propose RIATIG, a reliable and imperceptible adversarial attack against text-to-image models via inconspicuous examples. By formulating the example crafting as an optimization process and solving it using a genetic-based method, our proposed attack can generate imperceptible prompts for text-to-image generation models in a reliable way. Evaluation of six popular text-to-image generation models demonstrates the efficiency and stealthiness of our attack in both white-box and black-box settings. To allow the community to build on top of our findings, we've made the artifacts available.

We address the challenge of recovering an underlying scene geometry and colors from a sparse set of RGBD view observations. In this work, we present a new solution termed RGBD2 that sequentially generates novel RGBD views along a camera trajectory, and the scene geometry is simply the fusion result of these views. More specifically, we maintain an intermediate surface mesh used for rendering new RGBD views, which subsequently becomes complete by an inpainting network; each rendered RGBD view is later back-projected as a partial surface and is supplemented into the intermediate mesh. The use of intermediate mesh and camera projection helps solve the tough problem of multi-view inconsistency. We practically implement the RGBD inpainting network as a versatile RGBD diffusion model, which is previously used for 2D generative modeling; we make a modification to its reverse diffusion process to enable our use. We evaluate our approach on the task of 3D scene synthesis from sparse RGBD inputs; extensive experiments on the ScanNet dataset demonstrate the superiority of our approach over existing ones. Project page: https://jblei.site/proj/rgbd-diffusion.

Most neural networks for computer vision are designed to infer using RGB images. However, these RGB images are commonly encoded in JPEG before saving to disk; decoding them imposes an unavoidable overhead for RGB networks. Instead, our work focuses on training Vision Transformers (ViT) directly from the encoded features of JPEG. This way, we can avoid most of the decoding overhead, accelerating data load. Existing works have studied this aspect but they focus on CNNs. Due to how these encoded features are structured, CNNs require heavy modification to their architecture to accept such data. Here, we show that this is not the case for ViTs. In addition, we tackle data augmentation directly on these encoded features, which to our knowledge, has not been explored in-depth for training in this setting. With these two improvements -- ViT and data augmentation -- we show that our ViT-Ti model achieves up to 39.2% faster training and 17.9% faster inference with no accuracy loss compared to the RGB counterpart.

Reconstructing dynamic 3D garment surfaces with open boundaries from monocular videos is an important problem as it provides a practical and low-cost solution for clothes digitization. Recent neural rendering methods achieve high-quality dynamic clothed human reconstruction results from monocular video, but these methods cannot separate the garment surface from the body. Moreover, despite existing garment reconstruction methods based on feature curve representation demonstrating impressive results for garment reconstruction from a single image, they struggle to generate temporally consistent surfaces for the video input. To address the above limitations, in this paper, we formulate this task as an optimization problem of 3D garment feature curves and surface reconstruction from monocular video. We introduce a novel approach, called REC-MV to jointly optimize the explicit feature curves and the implicit signed distance field (SDF) of the garments. Then the open garment meshes can be extracted via garment template registration in the canonical space. Experiments on multiple casually captured datasets show that our approach outperforms existing methods and can produce high-quality dynamic garment surfaces.

Contrastive Language-Image Pre-training (CLIP) is attracting increasing attention for its impressive zero-shot recognition performance on different down-stream tasks. However, training CLIP is data-hungry and requires lots of image-text pairs to memorize various semantic concepts. In this paper, we propose a novel and efficient framework: Retrieval Augmented Contrastive Language-Image Pre-training (RA-CLIP) to augment embeddings by online retrieval. Specifically, we sample part of image-text data as a hold-out reference set. Given an input image, relevant image-text pairs are retrieved from the reference set to enrich the representation of input image. This process can be considered as an open-book exam: with the reference set as a cheat sheet, the proposed method doesn't need to memorize all visual concepts in the training data. It explores how to recognize visual concepts by exploiting correspondence between images and texts in the cheat sheet. The proposed RA-CLIP implements this idea and comprehensive experiments are conducted to show how RA-CLIP works. Performances on 10 image classification datasets and 2 object detection datasets show that RA-CLIP outperforms vanilla CLIP baseline by a large margin on zero-shot image classification task (+12.7%), linear probe image classification task (+6.9%) and zero-shot ROI classification task (+2.8%).

Recently, video moment retrieval and highlight detection (MR/HD) are being spotlighted as the demand for video understanding is drastically increased. The key objective of MR/HD is to localize the moment and estimate clip-wise accordance level, i.e., saliency score, to the given text query. Although the recent transformer-based models brought some advances, we found that these methods do not fully exploit the information of a given query. For example, the relevance between text query and video contents is sometimes neglected when predicting the moment and its saliency. To tackle this issue, we introduce Query-Dependent DETR (QD-DETR), a detection transformer tailored for MR/HD. As we observe the insignificant role of a given query in transformer architectures, our encoding module starts with cross-attention layers to explicitly inject the context of text query into video representation. Then, to enhance the model's capability of exploiting the query information, we manipulate the video-query pairs to produce irrelevant pairs. Such negative (irrelevant) video-query pairs are trained to yield low saliency scores, which in turn, encourages the model to estimate precise accordance between query-video pairs. Lastly, we present an input-adaptive saliency predictor which adaptively defines the criterion of saliency scores for the given video-query pairs. Our extensive studies verify the importance of building the query-dependent representation for MR/HD. Specifically, QD-DETR outperforms state-of-the-art methods on QVHighlights, TVSum, and Charades-STA datasets. Codes are available at github.com/wjun0830/QD-DETR.

Predicting the future trajectories of surrounding agents is essential for autonomous vehicles to operate safely. This paper presents QCNet, a modeling framework toward pushing the boundaries of trajectory prediction. First, we identify that the agent-centric modeling scheme used by existing approaches requires re-normalizing and re-encoding the input whenever the observation window slides forward, leading to redundant computations during online prediction. To overcome this limitation and achieve faster inference, we introduce a query-centric paradigm for scene encoding, which enables the reuse of past computations by learning representations independent of the global spacetime coordinate system. Sharing the invariant scene features among all target agents further allows the parallelism of multi-agent trajectory decoding. Second, even given rich encodings of the scene, existing decoding strategies struggle to capture the multimodality inherent in agents' future behavior, especially when the prediction horizon is long. To tackle this challenge, we first employ anchor-free queries to generate trajectory proposals in a recurrent fashion, which allows the model to utilize different scene contexts when decoding waypoints at different horizons. A refinement module then takes the trajectory proposals as anchors and leverages anchor-based queries to refine the trajectories further. By supplying adaptive and high-quality anchors to the refinement module, our query-based decoder can better deal with the multimodality in the output of trajectory prediction. Our approach ranks 1st on Argoverse 1 and Argoverse 2 motion forecasting benchmarks, outperforming all methods on all main metrics by a large margin. Meanwhile, our model can achieve streaming scene encoding and parallel multi-agent decoding thanks to the query-centric design ethos.

Hyperspectral image (HSI) classification aims at assigning a unique label for every pixel to identify categories of different land covers. Existing deep learning models for HSIs are usually performed in a traditional learning paradigm. Being emerging machines, quantum computers are limited in the noisy intermediate-scale quantum (NISQ) era. The quantum theory offers a new paradigm for designing deep learning models. Motivated by the quantum circuit (QC) model, we propose a quantum-inspired spectral-spatial network (QSSN) for HSI feature extraction. The proposed QSSN consists of a phase-prediction module (PPM) and a measurement-like fusion module (MFM) inspired from quantum theory to dynamically fuse spectral and spatial information. Specifically, QSSN uses a quantum representation to represent an HSI cuboid and extracts joint spectral-spatial features using MFM. An HSI cuboid and its phases predicted by PPM are used in the quantum representation. Using QSSN as the building block, we propose an end-to-end quantum-inspired spectral-spatial pyramid network (QSSPN) for HSI feature extraction and classification. In this pyramid framework, QSSPN progressively learns feature representations by cascading QSSN blocks and performs classification with a softmax classifier. It is the first attempt to introduce quantum theory in HSI processing model design. Substantial experiments are conducted on three HSI datasets to verify the superiority of the proposed QSSPN framework over the state-of-the-art methods.

While 3D-based GAN techniques have been successfully applied to render photo-realistic 3D images with a variety of attributes while preserving view consistency, there has been little research on how to fine-control 3D images without limiting to a specific category of objects of their properties. To fill such research gap, we propose a novel image manipulation model of 3D-based GAN representations for a fine-grained control of specific custom attributes. By extending the latest 3D-based GAN models (e.g., EG3D), our user-friendly quantitative manipulation model enables a fine yet normalized control of 3D manipulation of multi-attribute quantities while achieving view consistency. We validate the effectiveness of our proposed technique both qualitatively and quantitatively through various experiments.

The mechanism of existing style transfer algorithms is by minimizing a hybrid loss function to push the generated image toward high similarities in both content and style. However, this type of approach cannot guarantee visual fidelity, i.e., the generated artworks should be indistinguishable from real ones. In this paper, we devise a new style transfer framework called QuantArt for high visual-fidelity stylization. QuantArt pushes the latent representation of the generated artwork toward the centroids of the real artwork distribution with vector quantization. By fusing the quantized and continuous latent representations, QuantArt allows flexible control over the generated artworks in terms of content preservation, style similarity, and visual fidelity. Experiments on various style transfer settings show that our QuantArt framework achieves significantly higher visual fidelity compared with the existing style transfer methods.

Blind image quality assessment (BIQA) aims to automatically evaluate the perceived quality of a single image, whose performance has been improved by deep learning-based methods in recent years. However, the paucity of labeled data somewhat restrains deep learning-based BIQA methods from unleashing their full potential. In this paper, we propose to solve the problem by a pretext task customized for BIQA in a self-supervised learning manner, which enables learning representations from orders of magnitude more data. To constrain the learning process, we propose a quality-aware contrastive loss based on a simple assumption: the quality of patches from a distorted image should be similar, but vary from patches from the same image with different degradations and patches from different images. Further, we improve the existing degradation process and form a degradation space with the size of roughly 2x10^7. After pre-trained on ImageNet using our method, models are more sensitive to image quality and perform significantly better on downstream BIQA tasks. Experimental results show that our method obtains remarkable improvements on popular BIQA datasets.

Finetuning a large vision language model (VLM) on a target dataset after large scale pretraining is a dominant paradigm in visual question answering (VQA). Datasets for specialized tasks such as knowledge-based VQA or VQA in non natural-image domains are orders of magnitude smaller than those for general-purpose VQA. While collecting additional labels for specialized tasks or domains can be challenging, unlabeled images are often available. We introduce SelTDA (Self-Taught Data Augmentation), a strategy for finetuning large VLMs on small-scale VQA datasets. SelTDA uses the VLM and target dataset to build a teacher model that can generate question-answer pseudolabels directly conditioned on an image alone, allowing us to pseudolabel unlabeled images. SelTDA then finetunes the initial VLM on the original dataset augmented with freshly pseudolabeled images. We describe a series of experiments showing that our self-taught data augmentation increases robustness to adversarially searched questions, counterfactual examples, and rephrasings, it improves domain generalization, and results in greater retention of numerical reasoning skills. The proposed strategy requires no additional annotations or architectural modifications, and is compatible with any modern encoder-decoder multimodal transformer. Code available at https://github.com/codezakh/SelTDA

During industrial processing, unforeseen defects may arise in products due to uncontrollable factors. Although unsupervised methods have been successful in defect localization, the usual use of pre-trained models results in low-resolution outputs, which damages visual performance. To address this issue, we propose PyramidFlow, the first fully normalizing flow method without pre-trained models that enables high-resolution defect localization. Specifically, we propose a latent template-based defect contrastive localization paradigm to reduce intra-class variance, as the pre-trained models do. In addition, PyramidFlow utilizes pyramid-like normalizing flows for multi-scale fusing and volume normalization to help generalization. Our comprehensive studies on MVTecAD demonstrate the proposed method outperforms the comparable algorithms that do not use external priors, even achieving state-of-the-art performance in more challenging BTAD scenarios.

Deep learning has had remarkable success in robotic perception, but its data-centric nature suffers when it comes to generalizing to ever-changing environments. By contrast, physics-based optimization generalizes better, but it does not perform as well in complicated tasks due to the lack of high-level semantic information and reliance on manual parametric tuning. To take advantage of these two complementary worlds, we present PyPose: a robotics-oriented, PyTorch-based library that combines deep perceptual models with physics-based optimization. PyPose's architecture is tidy and well-organized, it has an imperative style interface and is efficient and user-friendly, making it easy to integrate into real-world robotic applications. Besides, it supports parallel computing of any order gradients of Lie groups and Lie algebras and 2nd-order optimizers, such as trust region methods. Experiments show that PyPose achieves more than 10x speedup in computation compared to the state-of-the-art libraries. To boost future research, we provide concrete examples for several fields of robot learning, including SLAM, planning, control, and inertial navigation.