ICCV 2023 Papers

Domain generalization (DG) aims to learn a robust model from source domains that generalize well on unseen target domains. Recent studies focus on generating novel domain samples or features to diversify distributions complementary to source domains. Yet, these approaches can hardly deal with the restriction that the samples synthesized from various domains can cause semantic distortion. In this paper, we propose a Cross Contrasting Feature Perturbation (CCFP) framework to simulate domain shift by generating perturbed features in the latent space while regularizing the model prediction against domain shift. Different from the previous fixed synthesizing strategy, we design modules with learnable feature perturbations and semantic consistency constraints. In contrast to prior work, our method does not use any generative-based models or domain labels. We conduct extensive experiments on a standard DomainBed benchmark with a strict evaluation protocol for a fair comparison. Comprehensive experiments show that our method outperforms the previous state-of-the-art, and quantitative analyses illustrate that our approach can alleviate the domain shift problem in out-of-distribution (OOD) scenarios.

Despite impressive performance for high-level downstream tasks, self-supervised pre-training methods have not yet fully delivered on dense geometric vision tasks such as stereo matching or optical flow. The application of self-supervised concepts, such as instance discrimination or masked image modeling, to geometric tasks is an active area of research. In this work, we build on the recent cross-view completion framework, a variation of masked image modeling that leverages a second view from the same scene which makes it well suited for binocular downstream tasks. The applicability of this concept has so far been limited in at least two ways: (a) by the difficulty of collecting real-world image pairs -- in practice only synthetic data have been used -- and (b) by the lack of generalization of vanilla transformers to dense downstream tasks for which relative position is more meaningful than absolute position. We explore three avenues of improvement. First, we introduce a method to collect suitable real-world image pairs at large scale. Second, we experiment with relative positional embeddings and show that they enable vision transformers to perform substantially better. Third, we scale up vision transformer based cross-completion architectures, which is made possible by the use of large amounts of data. With these improvements, we show for the first time that state-of-the-art results on stereo matching and optical flow can be reached without using any classical task-specific techniques like correlation volume, iterative estimation, image warping or multi-scale reasoning, thus paving the way towards universal vision models.

In this paper, we propose a novel method for single-view 3D style transfer that generates a unique 3D object with both shape and texture transfer. Our focus lies primarily on birds, a popular subject in 3D reconstruction, for which no existing single-view 3D transfer methods have been developed. The method we propose seeks to generate a 3D mesh shape and texture of a bird from two single-view images. To achieve this, we introduce a novel shape transfer generator that comprises a dual residual gated network (DRGNet), and a multi-layer perceptron (MLP). DRGNet extracts the features of source and target images using a shared coordinate gate unit, while the MLP generates spatial coordinates for building a 3D mesh. We also introduce a semantic UV texture transfer module that implements textural style transfer using semantic UV segmentation, which ensures consistency in the semantic meaning of the transferred regions. This module can be widely adapted to many existing approaches. Finally, our method constructs a novel 3D bird using a differentiable renderer. Experimental results on the CUB dataset verify that our method achieves state-of-the-art performance on the single-view 3D style transfer task. Code is available at https://github.com/wrk226/creative_birds.

Crowd counting has recently attracted significant attention in the field of computer vision due to its wide applications to image understanding. Numerous methods have been proposed and achieved state-of-the-art performance for real-world tasks. However, existing approaches do not perform well under adverse weather such as haze, rain, and snow since the visual appearances of crowds in such scenes are drastically different from those images in clear weather of typical datasets. In this paper, we propose a method for robust crowd counting in adverse weather scenarios. Instead of using a two-stage approach that involves image restoration and crowd counting modules, our model learns effective features and adaptive queries to account for large appearance variations. With these weather queries, the proposed model can learn the weather information according to the degradation of the input image and optimize with the crowd counting module simultaneously. Experimental results show that the proposed algorithm is effective in counting crowds under different weather types on benchmark datasets. The source code and trained models will be made available to the public.

Explaining deep models in a human-understandable way has been explored by many works that mostly explain why an input causes a corresponding prediction (ie., Why P?). However, seldom they could handle those more complex causal questions like "why P rather than Q?" and "why one is P while another is Q?", which would better help humans understand the behavior of deep models. Considering the insufficient study on such complex causal questions, we make the first attempt to explain different causal questions by contrastive explanations in a unified framework, ie., Counterfactual Contrastive Explanation (CCE), which visually and intuitively explains the aforementioned questions via a novel positive-negative saliency-based explanation scheme. More specifically, we propose a content-aware counterfactual perturbing algorithm to stimulate contrastive examples, from which a pair of positive and negative saliency maps could be derived to contrastively explain why P (positive class) rather than Q (negative class). Beyond existing works, our counterfactual perturbation meets the principles of validity, sparsity, and data distribution closeness at the same time. In addition, by slightly adjusting the objective of perturbation, our framework can adapt to different causal questions. Extensive experimental evaluation demonstrates the effectiveness and superior performance of the proposed CCE on different benchmark metrics for interpretability, including Sanity Check, Class Deviation Score and Insertion-Deletion tests. A user study is conducted and the results show that user confidence is increasing significantly when presented with CCE compared to standard saliency map baselines.

The inability of DNNs to explain their black-box behavior has led to a recent surge of explainability methods. However, there are growing concerns that these explainability methods are not robust and trustworthy. In this work, we perform the first robustness analysis of Neuron Explanation Methods under a unified pipeline and show that these explanations can be significantly corrupted by random noises and well-designed perturbations added to their probing data. We find that even adding small random noise with a standard deviation of 0.02 can already change the assigned concepts of up to 28% neurons in the deeper layers. Furthermore, we devise a novel corruption algorithm and show that our algorithm can manipulate the explanation of more than 80% neurons by poisoning less than 10% of probing data. This raises the concern of trusting Neuron Explanation Methods in real-life safety and fairness critical applications.

Neural Radiance Fields (NeRF) have the potential to be a major representation of media. Since training a NeRF has never been an easy task, the protection of its model copyright should be a priority. In this paper, by analyzing the pros and cons of possible copyright protection solutions, we propose to protect the copyright of NeRF models by replacing the original color representation in NeRF with a watermarked color representation. Then, a distortion-resistant rendering scheme is designed to guarantee robust message extraction in 2D renderings of NeRF. Our proposed method can directly protect the copyright of NeRF models while maintaining high rendering quality and bit accuracy when compared among optional solutions.

Multi-person 3D mesh recovery from videos is a critical first step towards automatic perception of group behavior in virtual reality, physical therapy and beyond. However, existing approaches rely on multi-stage paradigms, where the person detection and tracking stages are performed in a multi-person setting, while temporal dynamics are only modeled for one person at a time. Consequently, their performance is severely limited by the lack of inter-person interactions in the spatial-temporal mesh recovery, as well as by detection and tracking defects. To address these challenges, we propose the Coordinate transFormer (CoordFormer) that directly models multi-person spatial-temporal relations and simultaneously performs multi-mesh recovery in an end-to-end manner. Instead of partitioning the feature map into coarse-scale patch-wise tokens, CoordFormer leverages a novel Coordinate-Aware Attention to preserve pixel-level spatial-temporal coordinate information. Additionally, we propose a simple, yet effective Body Center Attention mechanism to fuse position information. Extensive experiments on the 3DPW dataset demonstrate that CoordFormer significantly improves the state-of-the-art, outperforming the previously best results by 4.2%, 8.8% and 4.7% according to the MPJPE, PAMPJPE, and PVE metrics, respectively, while being 40% faster than recent video-based approaches. The released code can be found at https://github.com/Li-Hao-yuan/CoordFormer.

In recent years, huge progress has been made on learn- ing neural implicit representations from multi-view images for 3D reconstruction. As an additional input complement- ing coordinates, using sinusoidal functions as positional encodings plays a key role in revealing high frequency de- tails with coordinate-based neural networks. However, high frequency positional encodings make the optimization un- stable, which results in noisy reconstructions and artifacts in empty space. To resolve this issue in a general sense, we introduce to learn neural implicit representations with quantized coordinates, which reduces the uncertainty and ambiguity in the field during optimization. Instead of con- tinuous coordinates, we discretize continuous coordinates into discrete coordinates using nearest interpolation among quantized coordinates which are obtained by discretizing the field in an extremely high resolution. We use discrete coordinates and their positional encodings to learn implicit functions through volume rendering. This significantly re- duces the variations in the sample space, and triggers more multi-view consistency constraints on intersections of rays from different views, which enables to infer implicit function in a more effective way. Our quantized coordinates do not bring any computational burden, and can seamlessly work upon the latest methods. Our evaluations under the widely used benchmarks show our superiority over the state-of-the- art. Our code is available at https://github.com/ MachinePerceptionLab/CQ-NIR.

In computer vision, 2D convolution is arguably the most important operation performed by a ConvNet. Unsurprisingly, it has been the focus of intense software and hardware optimization and enjoys highly efficient implementations. In this work, we ask an intriguing question: can we make a ConvNet work without 2D convolutions? Surprisingly, we find that the answer is yes --- we show that a ConvNet consisting entirely of 1D convolutions can do just as well as 2D on ImageNet classification. Specifically, we find that one key ingredient to a high-performing 1D ConvNet is oriented 1D kernels: 1D kernels that are oriented not just horizontally or vertically, but also at other angles. Our experiments show that oriented 1D convolutions can not only replace 2D convolutions but also augment existing architectures with large kernels, leading to improved accuracy with minimal FLOPs increase. A key contribution of this work is a highly-optimized custom CUDA implementation of oriented 1D kernels, specialized to the depthwise convolution setting. Our benchmarks demonstrate that our custom CUDA implementation almost perfectly realizes the theoretical advantage of 1D convolution: it is faster than a native horizontal convolution for any arbitrary angle. Code is available at https://github.com/princeton-vl/Oriented1D.

We describe a method to parse a complex, cluttered indoor scene into primitives which offer a parsimonious abstraction of scene structure. Our primitives are simple convexes. Our method uses a learned regression procedure to parse a scene into a fixed number of convexes from RGBD input, and can optionally accept segmentations to improve the decomposition. The result is then polished with a descent method which adjusts the convexes to produce a very good fit, and greedily removes superfluous primitives. Because the entire scene is parsed, we can evaluate using traditional depth, normal, and segmentation error metrics. Our evaluation procedure demonstrates that the error from our primitive representation is comparable to that of predicting depth from a single image.

Deep generative models have various content creation applications such as graphic design, e-commerce, and virtual try-on. However, current works mainly focus on synthesizing realistic visual outputs, often ignoring other sensory modalities, such as touch, which limits physical interaction with users. In this work, we leverage deep generative models to create a multi-sensory experience where users can touch and see the synthesized object when sliding their fingers on a haptic surface. The main challenges lie in the significant scale discrepancy between vision and touch sensing and the lack of explicit mapping from touch sensing data to a haptic rendering device. To bridge this gap, we collect high-resolution tactile data with a GelSight sensor and create a new visuotactile clothing dataset. We then develop a conditional generative model that synthesizes both visual and tactile outputs from a single sketch. We evaluate our method regarding image quality and tactile rendering accuracy. Finally, we introduce a pipeline to render high-quality visual and tactile outputs on an electroadhesion-based haptic device for an immersive experience, allowing for challenging materials and editable sketch inputs.

Controllable person image synthesis aims at rendering a source image based on user-specified changes in body pose or appearance. Prior art approaches leverage pixel-level denoising diffusion models conditioned on the coarse skeleton via cross-attention. This leads to two limitations: low efficiency and inaccurate condition information. To address both issues, a novel Pose-Constrained Latent Diffusion model (PoCoLD) is introduced. Rather than using the skeleton as a sparse pose representation, we exploit DensePose which offers much richer body structure information. To effectively capitalize DensePose at a low cost, we propose an efficient pose-constrained attention module that is capable of modeling the complex interplay between appearance and pose. Extensive experiments show that our PoCoLD outperforms the state-of-the-art competitors in image synthesis fidelity. Critically, it runs 2x faster and consumes 3.6x smaller memory than the latest diffusion-model-based alternative during inference.

Recent studies on face forgery detection have shown satisfactory performance for methods involved in training datasets, but are not ideal enough for unknown domains. This motivates many works to improve the generalization, but forgery-irrelevant information, such as image background and identity, still exists in different domain features and causes unexpected clustering, limiting the generalization. In this paper, we propose a controllable guide-space (GS) method to enhance the discrimination of different forgery domains, so as to increase the forgery relevance of features and thereby improve the generalization. The well-designed guide-space can simultaneously achieve both the proper separation of forgery domains and the large distance between real-forgery domains in an explicit and controllable manner. Moreover, for better discrimination, we use a decoupling module to weaken the interference of forgery-irrelevant correlations between domains. Furthermore, we make adjustments to the decision boundary manifold according to the clustering degree of the same domain features within the neighborhood. Extensive experiments in multiple in-domain and cross-domain settings confirm that our method can achieve state-of-the-art generalization.

The challenge in sourcing attribution for forgery faces has gained widespread attention due to the rapid development of generative techniques. While many recent works have taken essential steps on GAN-generated faces, more threatening attacks related to identity swapping or expression transferring are still overlooked. And the forgery traces hidden in unknown attacks from the open-world unlabeled faces still remain under-explored. To push the related frontier research, we introduce a new benchmark called Open-World DeepFake Attribution (OW-DFA), which aims to evaluate attribution performance against various types of fake faces under open-world scenarios. Meanwhile, we propose a novel framework named Contrastive Pseudo Learning (CPL) for the OW-DFA task through 1) introducing a Global-Local Voting module to guide the feature alignment of forged faces with different manipulated regions, 2) designing a Confidence-based Soft Pseudo-label strategy to mitigate the pseudo-noise caused by similar methods in unlabeled set. In addition, we extend the CPL framework with a multi-stage paradigm that leverages pre-train technique and iterative learning to further enhance traceability performance. Extensive experiments verify the superiority of our proposed method on the OW-DFA and also demonstrate the interpretability of deepfake attribution task and its impact on improving the security of deepfake detection area.

Standard unsupervised domain adaptation methods adapt models from a source to a target domain using labeled source data and unlabeled target data jointly. In model adaptation, on the other hand, access to the labeled source data is prohibited, i.e., only the source-trained model and unlabeled target data are available. We investigate normal-to-adverse condition model adaptation for semantic segmentation, whereby image-level correspondences are available in the target domain. The target set consists of unlabeled pairs of adverse- and normal-condition street images taken at GPS-matched locations. Our method--CMA--leverages such image pairs to learn condition-invariant features via contrastive learning. In particular, CMA encourages features in the embedding space to be grouped according to their condition-invariant semantic content and not according to the condition under which respective inputs are captured. To obtain accurate cross-domain semantic correspondences, we warp the normal image to the viewpoint of the adverse image and leverage warp-confidence scores to create robust, aggregated features. With this approach, we achieve state-of-the-art semantic segmentation performance for model adaptation on several normal-to-adverse adaptation benchmarks, such as ACDC and Dark Zurich. We also evaluate CMA on a newly procured adverse-condition generalization benchmark and report favorable results compared to standard unsupervised domain adaptation methods, despite the comparative handicap of CMA due to source data inaccessibility. Code is available at https://github.com/brdav/cma.

Though self-supervised contrastive learning (CL) has shown its potential to achieve state-of-the-art accuracy without any supervision, its behavior still remains under investigated by academia. Different from most previous work that understands CL from learning objectives, we focus on an unexplored yet natural aspect: the spatial inductive bias which seems to be implicitly exploited via data augmentations in CL. We design an experiment to study the reliance of CL on such spatial inductive bias, by destroying the global or local spatial structures of image with global or local patch shuffling, and comparing the performance drop between experiments on original and corrupted dataset to quantify the reliance of certain inductive bias. We also use the uniformity of feature space to further research on how CL-pre-trained model behave with the corrupted dataset. Our results and analysis show that CL has a much higher reliance on spatial inductive bias than SL, regardless of specific CL algorithm or backbones, opening a new direction for studying the behavior of CL.

We present Contrastive Feature Masking Vision Transformer (CFM-ViT) - an image-text pretraining methodology that achieves simultaneous learning of image- and region level representation for open-vocabulary object detection (OVD). Our approach combines the masked autoencoder (MAE) objective into the contrastive learning objective to improve the representation for localization tasks. Unlike standard MAE, we perform reconstruction in the joint image-text embedding space, rather than the pixel space as is customary with the classical MAE method, which causes the model to better learn region-level semantics. Moreover, we introduce Positional Embedding Dropout (PED) to address scale variation between image-text pretraining and detection finetuning by randomly dropping out the positional embeddings during pretraining. PED improves detection performance and enables the use of a frozen ViT backbone as a region classifier, preventing the forgetting of open-vocabulary knowledge during detection finetuning. On LVIS open-vocabulary detection benchmark, CFM-ViT achieves a state-of-the-art 33.9 APr, surpassing the best approach by 7.6 points and achieves better zero-shot detection transfer. Finally, CFM-ViT acquires strong image-level representation, outperforming the state of the art on 8 out of 12 metrics on zero-shot image-text retrieval benchmarks.

Self-supervised learning has attracted a lot of attention recently, which is able to learn powerful representations without any manual annotations. However, self-supervised learning needs to develop the ability to continuously learn to cope with a variety of real-world challenges, i.e., Continual Self-Supervised Learning (CSSL). Catastrophic forgetting is a notorious problem in CSSL, where the model tends to forget the learned knowledge. In practice, simple rehearsal or regularization will bring extra negative effects while alleviating catastrophic forgetting in CSSL, e.g., overfitting on the rehearsal samples or hindering the model from encoding fresh information. In order to address catastrophic forgetting without overfitting on the rehearsal samples, we propose Augmentation Stability Rehearsal (ASR) in this paper, which selects the most representative and discriminative samples by estimating the augmentation stability for rehearsal. Meanwhile, we design a matching strategy for ASR to dynamically update the rehearsal buffer. In addition, we further propose Contrastive Continuity on Augmentation Stability Rehearsal (C2ASR) based on ASR. We show that C2ASR is an upper bound of the Information Bottleneck (IB) principle, which suggests that C2ASR essentially preserves as much information shared among seen task streams as possible to prevent catastrophic forgetting and dismisses the redundant information between previous task streams and current task stream to free up the ability to encode fresh information. Our method obtains a great achievement compared with state-of-the-art CSSL methods on a variety of CSSL benchmarks.