ICCV 2023 Papers

While discriminative correlation filters (DCF)-based trackers prevail in UAV tracking for their favorable efficiency, lightweight convolutional neural network (CNN)-based trackers using filter pruning have also demonstrated remarkable efficiency and precision. However, the use of pure vision transformer models (ViTs) for UAV tracking remains unexplored, which is a surprising finding given that ViTs have been shown to produce better performance and greater efficiency than CNNs in image classification. In this paper, we propose an efficient ViT-based tracking framework, Aba-ViTrack, for UAV tracking. In our framework, feature learning and template-search coupling are integrated into an efficient one-stream ViT to avoid an extra heavy relation modeling module. The proposed Aba-ViT exploits an adaptive and background-aware token computation method to reduce inference time. This approach adaptively discards tokens based on learned halting probabilities, which a priori are higher for background tokens than target ones. Extensive experiments on six UAV tracking benchmarks demonstrate that the proposed Aba-ViTrack achieves state-of-the-art performance in UAV tracking. Code is available at https://github.com/xyyang317/Aba-ViTrack.

Mitochondria, as tiny structures within the cell, are of significant importance to study cell functions for biological and clinical analysis. And exploring how to automatically segment mitochondria in electron microscopy (EM) images has attracted increasing attention. However, most of existing methods struggle to adapt to different scales and appearances of the input due to the inherent limitations of the traditional CNN architecture. To mitigate these limitations, we propose a novel adaptive template transformer (ATFormer) for mitochondria segmentation. The proposed ATFormer model enjoys several merits. First, the designed structural template learning module can acquire appearance-adaptive templates of background, foreground and contour to sense the characteristics of different shapes of mitochondria. And we further adopt an optimal transport algorithm to enlarge the discrepancy among diverse templates to fully activate corresponding regions. Second, we introduce a hierarchical attention learning mechanism to absorb multi-level information for templates to be adaptive scale-aware classifiers for dense prediction. Extensive experimental results on three challenging benchmarks including MitoEM, Lucchi and NucMM-Z datasets demonstrate that our ATFormer performs favorably against state-of-the-art mitochondria segmentation methods.

Learning semantic segmentation requires pixel-wise annotations, which can be time-consuming and expensive. To reduce the annotation cost, we propose a superpixel-based active learning (AL) framework, which collects a dominant label per superpixel instead. To be specific, it consists of adaptive superpixel and sieving mechanisms, fully dedicated to AL. At each round of AL, we adaptively merge neighboring pixels of similar learned features into superpixels. We then query a selected subset of these superpixels using an acquisition function assuming no uniform superpixel size. This approach is more efficient than existing methods, which rely only on innate features such as RGB color and assume uniform superpixel sizes. Obtaining a dominant label per superpixel drastically reduces annotators' burden as it requires fewer clicks. However, it inevitably introduces noisy annotations due to mismatches between superpixel and ground truth segmentation. To address this issue, we further devise a sieving mechanism that identifies and excludes potentially noisy annotations from learning. Our experiments on both Cityscapes and PASCAL VOC datasets demonstrate the efficacy of adaptive superpixel and sieving mechanisms.

The success of deep learning models on structured data has generated significant interest in extending their application to non-Euclidean domains. In this work, we introduce a novel intrinsic operator suitable for representation learning on 3D meshes. Our operator is specifically tailored to adapt its behavior to the irregular structure of the underlying graph and effectively utilize its long-range dependencies, while at the same time ensuring computational efficiency and ease of optimization. In particular, inspired by the framework of Spiral Convolution, which extracts and transforms the vertices in the 3D mesh following a local spiral ordering, we propose a general operator that dynamically adjusts the length of the spiral trajectory and the parameters of the transformation for each processed vertex and mesh. Then, we use polyadic decomposition to factorize its dense weight tensor into a sequence of lighter linear layers that separately process features and vertices information, hence significantly reducing the computational complexity without introducing any stringent inductive biases. Notably, we leverage dynamic gating to achieve spatial adaptivity and induce global reasoning with constant time complexity benefitting from an efficient dynamic pooling mechanism based on Summed-Area-tables. Used as a drop-in replacement on existing architectures for shape correspondence our operator significantly improves the performance-efficiency trade-off, and in 3D shape generation with morphable models achieves state-of-the-art performance with a three-fold reduction in the number of parameters required. Project page: https://github.com/Fb2221/DFC

Most self-supervised methods for representation learning leverage a cross-view consistency objective i.e., they maximize the representation similarity of a given image's augmented views. Recent work NNCLR goes beyond the cross-view paradigm and uses positive pairs from different images obtained via nearest neighbor bootstrapping in a contrastive setting. We empirically show that as opposed to the contrastive learning setting which relies on negative samples, incorporating nearest neighbor bootstrapping in a self-distillation scheme can lead to a performance drop or even collapse. We scrutinize the reason for this unexpected behavior and provide a solution. We propose to adaptively bootstrap neighbors based on the estimated quality of the latent space. We report consistent improvements compared to the naive bootstrapping approach and the original baselines. Our approach leads to performance improvements for various self-distillation method/backbone combinations and standard downstream tasks. Our code is publicly available at https://github.com/tileb1/AdaSim.

Rotated object detection aims to identify and locate objects in images with arbitrary orientation. In this scenario, the oriented directions of objects vary considerably across different images, while multiple orientations of objects exist within an image. This intrinsic characteristic makes it challenging for standard backbone networks to extract high-quality features of these arbitrarily orientated objects. In this paper, we present Adaptive Rotated Convolution (ARC) module to handle the aforementioned challenges. In our ARC module, the convolution kernels rotate adaptively to extract object features with varying orientations in different images, and an efficient conditional computation mechanism is introduced to accommodate the large orientation variations of objects within an image. The two designs work seamlessly in rotated object detection problem. Moreover, ARC can conveniently serve as a plug-and-play module in various vision backbones to boost their representation ability to detect oriented objects accurately. Experiments on commonly used benchmarks (DOTA and HRSC2016) demonstrate that equipped with our proposed ARC module in the backbone network, the performance of multiple popular oriented object detectors is significantly improved (e.g. +3.03% mAP on Rotated RetinaNet and +4.16% on CFA). Combined with the highly competitive method Oriented R-CNN, the proposed approach achieves state-of-the-art performance on the DOTA dataset with 81.77% mAP. Code is available at https://github.com/LeapLabTHU/ARC.

We propose a new sampler for robust estimators that always selects the sample with the highest probability of consisting only of inliers. After every unsuccessful iteration, the inlier probabilities are updated in a principled way via a Bayesian approach. The probabilities obtained by the deep network are used as prior (so-called neural guidance) inside the sampler. Moreover, we introduce a new loss that exploits, in a geometrically justifiable manner, the orientation and scale that can be estimated for any type of feature, e.g., SIFT or SuperPoint, to estimate two-view geometry. The new loss helps to learn higher-order information about the underlying scene geometry. Benefiting from the new sampler and the proposed loss, we combine the neural guidance with the state-of-the-art MAGSAC++. Adaptive Reordering Sampler with Neurally Guided MAGSAC (ARS-MAGSAC) is superior to the state-of-the-art in terms of accuracy and run-time on the PhotoTourism and KITTI datasets for essential and fundamental matrix estimation. The code and trained models are available at https://github.com/weitong8591/ars_magsac.

Neural Radiance Fields have shown great potential to synthesize novel views with only a few discrete image observations of the world. However, the requirement of accurate camera parameters to learn scene representations limits its further application. In this paper, we present adaptive positional encoding (APE) for bundle-adjusting neural radiance fields to reconstruct the neural radiance fields from unknown camera poses (or even intrinsics). Inspired by Fourier series regression, we investigate its relationship with the positional encoding method and therefore propose APE where all frequency bands are trainable. Furthermore, we introduce period-activated multilayer perceptrons (PMLPs) to construct the implicit network for the high-order scene representations and fine-grain gradients during backpropagation. Experimental results on public datasets demonstrate that the proposed method with APE and PMLPs can outperform the state-of-the-art methods in accurate camera poses and high-fidelity view synthesis.

Recent works for face editing usually manipulate the latent space of StyleGAN via the linear semantic directions. However, they usually suffer from the entanglement of facial attributes, need to tune the optimal editing strength, and are limited to binary attributes with strong supervision signals. This paper proposes a novel adaptive nonlinear latent transformation for disentangled and conditional face editing, termed AdaTrans. Specifically, our AdaTrans divides the manipulation process into several finer steps; i.e., the direction and size at each step are conditioned on both the facial attributes and the latent codes. In this way, AdaTrans describes an adaptive nonlinear transformation trajectory to manipulate the faces into target attributes while keeping other attributes unchanged. Then, AdaTrans leverages a predefined density model to constrain the learned trajectory in the distribution of latent codes by maximizing the likelihood of transformed latent code. Moreover, we also propose a disentangled learning strategy under a mutual information framework to eliminate the entanglement among attributes, which can further relax the need for labeled data. Consequently, AdaTrans enables a controllable face editing with the advantages of disentanglement, flexibility with non-binary attributes, and high fidelity. Extensive experimental results on various facial attributes demonstrate the qualitative and quantitative effectiveness of the proposed AdaTrans over existing state-of-the-art methods, especially in the most challenging scenarios with a large age gap and few labeled examples. The source code is available at https://github.com/Hzzone/AdaTrans.

Deep learning based methods have become the de-facto standard for various computer vision tasks. Nevertheless, they have repeatedly shown their vulnerability to various form of input perturbations such as pixels modification, region anonymization, etc. which are closely related to the adversarial attacks. This research particularly addresses the case of image anonymization, which is significantly important to preserve privacy and hence to secure digitized form of personal information from being exposed and potentially misused by different services that have captured it for various purposes. However, applying anonymization causes the classifier to provide different class decisions before and after applying it and therefore reduces the classifier's reliability and usability. In order to achieve a robust solution to this problem we propose a novel anonymization procedure that allows the existing classifiers to become class decision invariant on the anonymized images without any modification requires to apply on the classification models. We conduct numerous experiments on the popular ImageNet benchmark as well as on a large scale industrial toll classification problem's dataset. Obtained results confirm the efficiency and effectiveness of the proposed method as it obtained 0% rate of class decision change for both datasets compared to 15.95% on ImageNet and 0.18% on toll dataset obtained by applying the naive anonymization approaches. Moreover, it has shown a great potential to be applied to similar problems from different domains.

Shadow detection methods rely on multi-scale contrast, especially global contrast, information to locate shadows correctly. However, we observe that the camera image signal processor (ISP) tends to preserve more local contrast information by sacrificing global contrast information during the raw-to-sRGB conversion process. This often causes existing methods to fail in scenes with high global contrast but low local contrast in shadow regions. In this paper, we propose a novel method to detect shadows from raw images. Our key idea is that instead of performing a many-to-one mapping like the ISP process, we can learn a many-to-many mapping from the high dynamic range raw images to the sRGB images of different illumination, which is able to preserve multi-scale contrast for accurate shadow detection. To this end, we first construct a new shadow dataset with 7000 raw images and shadow masks. We then propose a novel network, which includes a novel adaptive illumination mapping (AIM) module to project the input raw images into sRGB images of different intensity ranges and a shadow detection module to leverage the preserved multi-scale contrast information to detect shadows. To learn the shadow-aware adaptive illumination mapping process, we propose a novel feedback mechanism to guide the AIM during training. Experiments show that our method outperforms state-of-the-art shadow detectors. Code and dataset are available at https://github.com/jiayusun/SARA.

Recent vision transformers, large-kernel CNNs and MLPs have attained remarkable successes in broad vision tasks thanks to their effective information fusion in the global scope. However, their efficient deployments, especially on mobile devices, still suffer from noteworthy challenges due to the heavy computational costs of self-attention mechanisms, large kernels, or fully connected layers. In this work, we apply conventional convolution theorem to deep learning for addressing this and reveal that adaptive frequency filters can serve as efficient global token mixer. With this insight, we propose Adaptive Frequency Filtering (AFF) token mixer. This neural operator transfers a latent representation to the frequency domain via a Fourier transform and performs semantic-adaptive frequency filtering via an elementwise multiplication, which mathematically equals to a token mixing operation in the original latent space with a dynamic convolution kernel as large as the spatial resolution of this latent representation. We take AFF token mixers as primary neural operators to build a lightweight neural network, dubbed AFFNet. Extensive experiments demonstrate the effectiveness of our proposed AFF token mixer and show that AFFNet achieve superior accuracy and efficiency trade-offs compared to other lightweight network designs on broad visual tasks, including visual recognition and dense prediction tasks.

Model calibration usually requires optimizing some parameters (e.g., temperature) w.r.t an objective function like negative log-likelihood. This work uncovers a significant aspect often overlooked that the objective function is influenced by calibration set difficulty: the ratio of misclassified to correctly classified samples. If a test set has a drastically different difficulty level from the calibration set, a phenomenon out-of-distribution (OOD) data often exhibit: the optimal calibration parameters of the two datasets would be different, rendering an optimal calibrator on the calibration set suboptimal on the OOD test set and thus degraded calibration performance. With this knowledge, we propose a simple and effective method named adaptive calibrator ensemble (ACE) to calibrate OOD datasets whose difficulty is usually higher than the calibration set. Specifically, two calibration functions are trained, one for in-distribution data (low difficulty), and the other for severely OOD data (high difficulty). To achieve desirable calibration on a new OOD dataset, ACE uses an adaptive weighting method that strikes a balance between the two extreme functions. When plugged in, ACE generally improves the performance of a few state-of-the-art calibration schemes on a series of OOD benchmarks. Importantly, such improvement does not come at the cost of the in-distribution calibration performance. Project Website: https://github.com/insysgroup/Adaptive-Calibrators-Ensemble.git.

Compressive autoencoders (CAEs) play an important role in deep learning-based image compression, but large-scale CAEs are computationally expensive. We propose a framework with three techniques to enable efficient CAE-based image coding: 1) Spatially-adaptive convolution and normalization operators enable block-wise nonlinear transform to spend FLOPs unevenly across the image to be compressed, according to a transform capacity map. 2) Just-unpenalized model capacity (JUMC) optimizes the transform capacity of each CAE block via rate-distortion-complexity optimization, finding the optimal capacity for the source image content. 3) A lightweight routing agent model predicts the transform capacity map for the CAEs by approximating JUMC targets. By activating the best-sized sub-CAE inside the slimmable supernet, our approach achieves up to 40% computational speed-up with minimal BD-Rate increase, validating its ability to save computational resources in a content-aware manner.

Sparsely activated Mixture-of-Experts (MoE) is becoming a promising paradigm for multi-task learning (MTL). Instead of compressing multiple tasks' knowledge into a single model, MoE separates the parameter space and only utilizes the relevant model pieces given task type and its input, which provides stabilized MTL training and ultra-efficient inference. However, current MoE approaches adopt a fixed network capacity (e.g., two experts in usual) for all tasks. It potentially results in the over-fitting of simple tasks or the under-fitting of challenging scenarios, especially when tasks are significantly distinctive in their complexity. In this paper, we propose an adaptive MoE framework for multi-task vision recognition, dubbed AdaMV-MoE. Based on the training dynamics, it automatically determines the number of activated experts for each task, avoiding the laborious manual tuning of optimal model size. To validate our proposal, we benchmark it on ImageNet classification and COCO object detection & instance segmentation which are notoriously difficult to learn in concert, due to their discrepancy. Extensive experiments across a variety of vision transformers demonstrate a superior performance of AdaMV-MoE, compared to MTL with a shared backbone and the recent state-of-the-art (SoTA) MTL MoE approach. Codes are available online: https://github.com/google-research/google-research/tree/master/moe_mtl.

Voxel-based methods have achieved state-of-the-art performance for 3D object detection in autonomous driving. However, their significant computational and memory costs pose a challenge for their application to resource-constrained vehicles. One reason for this high resource consumption is the presence of a large number of redundant background points in Lidar point clouds, resulting in spatial redundancy in both 3D voxel and dense BEV map representations. To address this issue, we propose an adaptive inference framework called Ada3D, which focuses on exploiting the input-level spatial redundancy. Ada3D adaptively filters the redundant input, guided by a lightweight importance predictor and the unique properties of the Lidar point cloud. Additionally, we utilize the BEV features' intrinsic sparsity by introducing the Sparsity Preserving Batch Normalization. With Ada3D, we achieve 40% reduction for 3D voxels and decrease the density of 2D BEV feature maps from 100% to 20% without sacrificing accuracy. Ada3D reduces the model computational and memory cost by 5x, and achieves 1.52x/1.45x end-to-end GPU latency and 1.5x/4.5x GPU peak memory optimization for the 3D and 2D backbone respectively.

We present AdVerb, a novel audio-visual dereverberation framework that uses visual cues in addition to the reverberant sound to estimate clean audio. Although audio-only dereverberation is a well-studied problem, our approach incorporates the complementary visual modality to perform audio dereverberation. Given an image of the environment where the reverberated sound signal has been recorded, AdVerb employs a novel geometry-aware cross-modal transformer architecture that captures scene geometry and audio-visual cross-modal relationship to generate a complex ideal ratio mask, which, when applied to the reverberant audio predicts the clean sound. The effectiveness of our method is demonstrated through extensive quantitative and qualitative evaluations. Our approach significantly outperforms traditional audio-only and audio-visual baselines on three downstream tasks: speech enhancement, speech recognition, and speaker verification, with relative improvements in the range of 18% - 82% on the LibriSpeech test-clean set. We also achieve highly satisfactory RT60 error scores on the AVSpeech dataset.

While NeRF-based human representations have shown impressive novel view synthesis results, most methods still rely on a large number of images / views for training. In this work, we propose a novel animatable NeRF called ActorsNeRF. It is first pre-trained on diverse human subjects, and then adapted with few-shot monocular video frames for a new actor with unseen poses. Building on previous generalizable NeRFs with parameter sharing using a ConvNet encoder, ActorsNeRF further adopts two human priors to capture the large human appearance, shape, and pose variations. Specifically, in the encoded feature space, we will first align different human subjects in a category-level canonical space, and then align the same human from different frames in an instance-level canonical space for rendering. We quantitatively and qualitatively demonstrate that ActorsNeRF significantly outperforms the existing state-of-the-art on few-shot generalization to new people and poses on multiple datasets.