论文检索

Xiaoxing Wang, Xiangxiang Chu, Yuda Fan, Zhexi Zhang, Bo Zhang, Xiaokang Yang, Junchi Yan

Albeit being a prevalent architecture searching approach, differentiable architecture search (DARTS) is largely hindered by its substantial memory cost since the entire supernet resides in the memory. This is where the single-path DARTS comes in, which only chooses a single-path submodel at each step. While being memory-friendly, it also comes with low computational costs. Nonetheless, we discover a critical issue of single-path DARTS that has not been primarily noticed. Namely, it also suffers from severe performance collapse since too many parameter-free operations like skip connections are derived, just like DARTS does. In this paper, we propose a new algorithm called RObustifying Memory-Efficient NAS (ROME) to give a cure. First, we disentangle the topology search from the operation search to make searching and evaluation consistent. We then adopt Gumbel-Top2 reparameterization and gradient accumulation to robustify the unwieldy bi-level optimization. We verify ROME extensively across 15 benchmarks to demonstrate its effectiveness and robustness.

Zi-Yuan Hu, Yanyang Li, Michael R. Lyu, Liwei Wang

As the model size of pre-trained language models (PLMs) grows rapidly, full fine-tuning becomes prohibitively expensive for model training and storage. In vision-and-language (VL), parameter-efficient tuning (PET) techniques are proposed to integrate modular modifications (e.g., Adapter) into encoder-decoder PLMs. By tuning a small set of trainable parameters, these techniques perform on par with full fine-tuning. However, excessive modular modifications and neglecting the unique abilities of the encoders and decoders can lead to performance degradation, while existing PET techniques (e.g., VL-Adapter) overlook these issues. In this paper, we propose a Vision-and-Language Parameter-Efficient Tuning (VL-PET) framework to impose effective control over modular modifications via a novel granularity-controlled mechanism. Considering different granularity-controlled matrices generated by this mechanism, a variety of model-agnostic VL-PET modules can be instantiated from our framework for better efficiency and effectiveness trade-offs. We further propose lightweight designs to enhance VL alignment and modeling for the encoders and maintain text generation for the decoders. Extensive experiments conducted on four image-text tasks and four video-text tasks demonstrate the efficiency, effectiveness, scalability and transferability of our VL-PET framework. In particular, our VL-PET-large significantly outperforms full fine-tuning by 2.39% (2.61%) and VL-Adapter by 2.92% (3.41%) with BART-base (T5-base) on image-text tasks, while utilizing fewer trainable parameters. Furthermore, we validate the enhanced effect of employing our VL-PET designs (e.g., granularity-controlled mechanism and lightweight designs) on existing PET techniques, enabling them to achieve significant performance improvements.

Yijun Qian, Jack Urbanek, Alexander G. Hauptmann, Jungdam Won

Given its wide applications, there is increasing focus on generating 3D human motions from textual descriptions. Differing from the majority of previous works, which regard actions as single entities and can only generate short sequences for simple motions, we propose EMS, an elaborative motion synthesis model conditioned on detailed natural language descriptions. It generates natural and smooth motion sequences for long and complicated actions by factorizing them into groups of atomic actions. Meanwhile, it understands atomic-action level attributes (e.g., motion direction, speed, and body parts) and enables users to generate sequences of unseen complex actions from unique sequences of known atomic actions with independent attribute settings and timings applied. We evaluate our method on the KIT Motion-Language and BABEL benchmarks, where it outperforms all previous state-of-the-art with noticeable margins.

Taeksoo Kim, Shunsuke Saito, Hanbyul Joo

Deep generative models have been recently extended to synthesizing 3D digital humans. However, previous approaches treat clothed humans as a single chunk of geometry without considering the compositionality of clothing and accessories. As a result, individual items cannot be naturally composed into novel identities, leading to limited expressiveness and controllability of generative 3D avatars. While several methods attempt to address this by leveraging synthetic data, the interaction between humans and objects is not authentic due to the domain gap, and manual asset creation is difficult to scale for a wide variety of objects. In this work, we present a novel framework for learning a compositional generative model of humans and objects (backpacks, coats, scarves, and more) from real-world 3D scans. Our compositional model is interaction-aware, meaning the spatial relationship between humans and objects, and the mutual shape change by physical contact is fully incorporated. The key challenge is that, since humans and objects are in contact, their 3D scans are merged into a single piece. To decompose them without manual annotations, we propose to leverage two sets of 3D scans of a single person with and without objects. Our approach learns to decompose objects and naturally compose them back into a generative human model in an unsupervised manner. Despite our simple setup requiring only the capture of a single subject with objects, our experiments demonstrate the strong generalization of our model by enabling the natural composition of objects to diverse identities in various poses and the composition of multiple objects, which is unseen in training data. The project page is available at https://taeksuu.github.io/ncho.

Yanjing Li, Sheng Xu, Mingbao Lin, Jihao Yin, Baochang Zhang, Xianbin Cao

In this paper, we focus on developing knowledge distillation (KD) for compact 3D detectors. We observe that off-the-shelf KD methods manifest their efficacy only when the teacher model and student counterpart share similar intermediate feature representations. This might explain why they are less effective in building extreme-compact 3D detectors where significant representation disparity arises due primarily to the intrinsic sparsity and irregularity in 3D point clouds. This paper presents a novel representation disparity-aware distillation (RDD) method to address the representation disparity issue and reduce performance gap between compact students and over-parameterized teachers. This is accomplished by building our RDD from an innovative perspective of information bottleneck (IB), which can effectively minimize the disparity of proposal region pairs from student and teacher in features and logits. Extensive experiments are performed to demonstrate the superiority of our RDD over existing KD methods. For example, our RDD increases mAP of CP-Voxel-S to 57.1% on nuScenes dataset, which even surpasses teacher performance while taking up only 42% FLOPs.

Man Zhou, Jie Huang, Naishan Zheng, Chongyi Li

The success of deep neural networks for pan-sharpening is commonly in a form of black box, lacking transparency and interpretability. To alleviate this issue, we propose a novel model-driven deep unfolding framework with image reasoning prior tailored for the pan-sharpening task. Different from existing unfolding solutions that deliver the proximal operator networks as the uncertain and vague priors, our framework is motivated by the content reasoning ability of masked autoencoders (MAE) with insightful designs. Specifically, the pre-trained MAE with spatial masking strategy, acting as intrinsic reasoning prior, is embedded into unfolding architecture. Meanwhile, the pre-trained MAE with spatial-spectral masking strategy is treated as the regularization term within loss function to constrain the spatial-spectral consistency. Such designs penetrate the image reasoning prior into deep unfolding networks while improving its interpretability and representation capability. The uniqueness of our framework is that the holistic learning process is explicitly integrated with the inherent physical mechanism underlying the pan-sharpening task. Extensive experiments on multiple satellite datasets demonstrate the superiority of our method over the existing state-of-the-art approaches.

Yadan Luo, Zhuoxiao Chen, Zhen Fang, Zheng Zhang, Mahsa Baktashmotlagh, Zi Huang

Achieving a reliable LiDAR-based object detector in autonomous driving is paramount, but its success hinges on obtaining large amounts of precise 3D annotations. Active learning (AL) seeks to mitigate the annotation burden through algorithms that use fewer labels and can attain performance comparable to fully supervised learning. Although AL has shown promise, current approaches prioritize the selection of unlabeled point clouds with high aleatoric and/or epistemic uncertainty, leading to the selection of more instances for labeling and reduced computational efficiency. In this paper, we resort to a novel kernel coding rate maximization (KECOR) strategy which aims to identify the most informative point clouds to acquire labels through the lens of information theory. Greedy search is applied to seek desired point clouds that can maximize the minimal number of bits required to encode the latent features. To determine the uniqueness and informativeness of the selected samples from the model perspective, we construct a proxy network of the 3D detector head and compute the outer product of Jacobians from all proxy layers to form the empirical neural tangent kernel (NTK) matrix. To accommodate both one-stage (i.e., SECOND) and two-stage detectors (i.e., PV-RCNN), we further incorporate the classification entropy maximization and well trade-off between detection performance and the total number of bounding boxes selected for annotation. Extensive experiments conducted on two 3D benchmarks and a 2D detection dataset evidence the superiority and versatility of the proposed approach. Our results show that approximately 44% box-level annotation costs and 26% computational time are reduced compared to the state-of-the-art AL method, without compromising detection performance.

Zijian Dong, Xu Chen, Jinlong Yang, Michael J. Black, Otmar Hilliges, Andreas Geiger

While progress in 2D generative models of human appearance has been rapid, many applications require 3D avatars that can be animated and rendered. Unfortunately, most existing methods for learning generative models of 3D humans with diverse shape and appearance require 3D training data, which is limited and expensive to acquire. The key to progress is hence to learn generative models of 3D avatars from abundant unstructured 2D image collections. However, learning realistic and complete 3D appearance and geometry in this under-constrained setting remains challenging, especially in the presence of loose clothing such as dresses. In this paper, we propose a new adversarial generative model of realistic 3D people from 2D images. Our method captures shape and deformation of the body and loose clothing by adopting a holistic 3D generator and integrating an efficient, flexible, articulation module. To improve realism, we train our model using multiple discriminators while also integrating geometric cues in the form of predicted 2D normal maps. We experimentally find that our method outperforms previous 3D- and articulation-aware methods in terms of geometry and appearance. We validate the effectiveness of our model and the importance of each component via systematic ablation studies.

Yangru Huang, Peixi Peng, Yifan Zhao, Yunpeng Zhai, Haoran Xu, Yonghong Tian

Efficient motion and appearance modeling are critical for vision-based Reinforcement Learning (RL). However, existing methods struggle to reconcile motion and appearance information within the state representations learned from a single observation encoder. To address the problem, we present Synergizing Interactive Motion-appearance Understanding (Simoun), a unified framework for vision-based RL. Given consecutive observation frames, Simoun deliberately and interactively learns both motion and appearance features through a dual-path network architecture. The learning process collaborates with a structural interactive module, which explores the latent motion-appearance structures from the two network paths to leverage their complementarity. To promote sample efficiency, we further design a consistency-guided curiosity module to encourage the exploration of under-learned observations. During training, the curiosity module provides intrinsic rewards according to the consistency of environmental temporal dynamics, which are deduced from both motion and appearance network paths. Experiments conducted on the DeepMind control suite and CARLA automatic driving benchmarks demonstrate the effectiveness of Simoun, where it performs favorably against state-of-the-art methods.

Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneeshan Sain, Subhadeep Koley, Tao Xiang, Yi-Zhe Song

This paper studies the problem of 2D sketch to 3D shape retrieval, but with a focus on democratising the process. We would like this democratisation to happen on two fronts: (i) to remove the need for large-scale specifically sourced 2D sketch and 3D shape datasets, and (ii) to remove restrictions on how well the user needs to sketch and from what viewpoint. The end result is a system that is trainable using existing datasets, and once trained allows users to sketch regardless of drawing skills and without restriction on view angle. We achieve all this via a clever use of pivoting, along with novel designs that injects 3D understanding of 2D sketches into the system. We perform pivoting on two existing datasets, each from a distant research domain to the other: 2D sketch and photo pairs from the sketch-based image retrieval field (SBIR), and 3D shapes from ShapeNet. It follows that the actual feature pivoting happens on photos from the former and 2D projections from the latter. Doing this already achieves most of our democratisation challenge -- the level of 2D sketch abstraction embedded in SBIR dataset offers demoralization on drawing quality, and the whole thing works without a specifically sourced 2D sketch and 3D model pair. To further achieve democratisation on sketching viewpoint, we "lift" 2D sketches to 3D space using Blind Perspective-n-Points (BPnP) that injects 3D-aware information into the sketch encoder. Results show ours achieves competitive performance compared with fully-supervised baselines, while meeting all set democratisation goals.

Fangzhou Lin, Yun Yue, Songlin Hou, Xuechu Yu, Yajun Xu, Kazunori D Yamada, Ziming Zhang

Chamfer distance (CD) is a standard metric to measure the shape dissimilarity between point clouds in point cloud completion, as well as a loss function for (deep) learning. However, it is well known that CD is vulnerable to outliers, leading to the drift towards suboptimal models. In contrast to the literature where most works address such issues in Euclidean space, we propose an extremely simple yet powerful metric for point cloud completion, namely Hyperbolic Chamfer Distance (HyperCD), that computes CD in hyperbolic space. In backpropagation, HyperCD consistently assigns higher weights to the matched point pairs with smaller Euclidean distances. In this way, good point matches are likely to be preserved while bad matches can be updated gradually, leading to better completion results. We demonstrate state-of-the-art performance on the benchmark datasets, i.e. PCN, ShapeNet-55, and ShapeNet-34, and show from visualization that HyperCD can significantly improve the surface smoothness. Code is available at: https://github.com/Zhang-VISLab.

Kan Wu, Houwen Peng, Zhenghong Zhou, Bin Xiao, Mengchen Liu, Lu Yuan, Hong Xuan, Michael Valenzuela, Xi (Stephen) Chen, Xinggang Wang 等

In this paper, we propose a novel cross-modal distillation method, called TinyCLIP, for large-scale language-image pre-trained models. The method introduces two core techniques: affinity mimicking and weight inheritance. Affinity mimicking explores the interaction between modalities during distillation, enabling student models to mimic teachers' behavior of learning cross-modal feature alignment in a visual-linguistic affinity space. Weight inheritance transmits the pre-trained weights from the teacher models to their student counterparts to improve distillation efficiency. Moreover, we extend the method into a multi-stage progressive distillation to mitigate the loss of informative weights during extreme compression. Comprehensive experiments demonstrate the efficacy of TinyCLIP, showing that it can reduce the size of the pre-trained CLIP ViT-B/32 by 50%, while maintaining comparable zero-shot performance. While aiming for comparable performance, distillation with weight inheritance can speed up the training by 1.4 - 7.8x compared to training from scratch. Moreover, our TinyCLIP ViT-8M/16, trained on YFCC-15M, achieves an impressive zero-shot top-1 accuracy of 41.1% on ImageNet, surpassing the original CLIP ViT-B/16 by 3.5% while utilizing only 8.9% parameters. Finally, we demonstrate the good transferability of TinyCLIP in various downstream tasks. Code and models will be open-sourced at aka.ms/tinyclip.

Tengda Han, Max Bain, Arsha Nagrani, Gul Varol, Weidi Xie, Andrew Zisserman

Audio Description (AD) is the task of generating descriptions of visual content, at suitable time intervals, for the benefit of visually impaired audiences. For movies, this presents notable challenges -- AD must occur only during existing pauses in dialogue, should refer to characters by name, and ought to aid understanding of the storyline as a whole. To this end, we develop a new model for automatically generating movie AD, given CLIP visual features of the frames, the cast list, and the temporal locations of the speech; addressing all three of the `who', `when', and `what' questions: (i) who -- we introduce a character bank consisting of the character's name, the actor that played the part, and a CLIP feature of their face, for the principal cast of each movie, and demonstrate how this can be used to improve naming in the generated AD; (ii) when -- we investigate several models for determining whether an AD should be generated for a time interval or not, based on the visual content of the interval and its neighbours; and (iii) what -- we implement a new vision-language model for this task, that can ingest the proposals from the character bank, whilst conditioning on the visual features using cross-attention, and demonstrate how this improves over previous architectures for AD text generation in an apples-to-apples comparison.

Xinyi Zhang, Naiqi Li, Jiawei Li, Tao Dai, Yong Jiang, Shu-Tao Xia

Unsupervised surface anomaly detection aims at discovering and localizing anomalous patterns using only anomaly-free training samples. Reconstruction-based models are among the most popular and successful methods, which rely on the assumption that anomaly regions are more difficult to reconstruct. However, there are three major challenges to the practical application of this approach: 1) the reconstruction quality needs to be further improved since it has a great impact on the final result, especially for images with structural changes; 2) it is observed that for many neural networks, the anomalies can also be well reconstructed, which severely violates the underlying assumption; 3) since reconstruction is an ill-conditioned problem, a test instance may correspond to multiple normal patterns, but most current reconstruction-based methods have ignored this critical fact. In this paper, we propose DiffAD, a method for unsupervised anomaly detection based on the latent diffusion model, inspired by its ability to generate high-quality and diverse images. We further propose noisy condition embedding and interpolated channels to address the aforementioned challenges in the general reconstruction-based pipeline. Extensive experiments show that our method achieves state-of-the-art performance on the challenging MVTec dataset, especially in localization accuracy.

Karsten Roth, Jae Myung Kim, A. Sophia Koepke, Oriol Vinyals, Cordelia Schmid, Zeynep Akata

The visual classification performance of vision-language models such as CLIP has been shown to benefit from additional semantic knowledge from large language models (LLMs) such as GPT-3. In particular, averaging over LLM-generated class descriptors, e.g. "waffle, which has a round shape", can notably improve generalization performance. In this work, we critically study this behavior and propose WaffleCLIP, a framework for zero-shot visual classification which simply replaces LLM-generated descriptors with random character and word descriptors. Without querying external models, we achieve comparable performance gains on a large number of visual classification tasks. This allows WaffleCLIP to both serve as a low-cost alternative, as well as a sanity check for any future LLM-based vision-language model extensions. We conduct an extensive experimental study on the impact and shortcomings of additional semantics introduced with LLM-generated descriptors, and showcase how - if available - semantic context is better leveraged by querying LLMs for high-level concepts, which we show can be done to jointly resolve potential class name ambiguities. Code is available here: https://github.com/ExplainableML/WaffleCLIP.

Yizhong Pan, Xiao Liu, Xiangyu Liao, Yuanzhouhan Cao, Chao Ren

With sufficient paired training samples, the supervised deep learning methods have attracted much attention in image denoising because of their superior performance. However, it is still very challenging to widely utilize the supervised methods in real cases due to the lack of paired noisy-clean images. Meanwhile, most self-supervised denoising methods are ineffective as well when applied to the real-world denoising tasks because of their strict assumptions in applications. For example, as a typical method for self-supervised denoising, the original blind spot network (BSN) assumes that the noise is pixel-wise independent, which is much different from the real cases. To solve this problem, we propose a novel self-supervised real image denoising framework named Sampling Difference As Perturbation (SDAP) based on Random Sub-samples Generation (RSG) with a cyclic sample difference loss. Specifically, we dig deeper into the properties of BSN to make it more suitable for real noise. Surprisingly, we find that adding an appropriate perturbation to the training images can effectively improve the performance of BSN. Further, we propose that the sampling difference can be considered as perturbation to achieve better results. Finally we propose a new BSN framework in combination with our RSG strategy. The results show that it significantly outperforms other state-of-the-art self-supervised denoising methods on real-world datasets. The code is available at https://github.com/p1y2z3/SDAP.

Songhua Liu, Xinchao Wang

Dataset distillation aims at a small synthetic dataset to mimic the training performance on neural networks of a given large dataset. Existing approaches heavily rely on an iterative optimization to update synthetic data and multiple forward-backward passes over thousands of neural network spaces, which introduce significant overhead for computation and are inconvenient in scenarios requiring high efficiency. In this paper, we focus on few-shot dataset distillation, where a distilled dataset is synthesized with only a few or even a single network. To this end, we introduce the notion of distillation space, such that synthetic data optimized only in this specific space can achieve the effect of those optimized through numerous neural networks, with dramatically accelerated training and reduced computational cost. To learn such a distillation space, we first formulate the problem as a quad-level optimization framework and propose a bi-level algorithm. Nevertheless, the algorithm in its original form has a large memory footprint in practice due to the back-propagation through an unrolled computational graph. We then convert the problem of learning the distillation space to a first-order one based on image translation. Specifically, the synthetic images are optimized in an arbitrary but fixed neural space and then translated to those in the targeted distillation space. We pre-train the translator on some large datasets like ImageNet so that it requires only a limited number of adaptation steps on the target dataset. Extensive experiments demonstrate that the translator after pre-training and a limited number of adaptation steps achieves comparable distillation performance with state of the arts, with 15x acceleration. It also exerts satisfactory generalization performance across different datasets, storage budgets, and numbers of classes.

Zhuchen Shao, Yifeng Wang, Yang Chen, Hao Bian, Shaohui Liu, Haoqian Wang, Yongbing Zhang

Gigapixel Whole Slide Images (WSIs) aided patient diagnosis and prognosis analysis are promising directions in computational pathology. However, limited by expensive and time-consuming annotation costs, WSIs usually only have weak annotations, including 1) WSI-level Annotations (WA) and 2) Limited Patch-level Annotations (LPA). Currently, Multiple Instance Learning (MIL) often exploits WA, while LPA usually assign pseudo-labels for unlabeled data. Intuitively, pseudo-labels can serve as a practical guide for MIL, but the unreliable prediction caused by LPA inevitably introduces noise. Furthermore, WA-supervised MIL training inevitably suffers from the semantical unalignment between instances and bag-level labels. To address these problems, we design a framework called Learning from Noisy Pseudo Labels for promoting Multiple Instance Learning (LNPL-MIL), which considers both types of weak annotation. In MIL, we propose a Transformer aware of instance Order and Distribution (TOD-MIL) that strengthens instances correlation and weakens semantical unalignment in the bag. We validate our LNPL-MIL on Tumor Diagnosis and Survival Prediction, achieving state-of-the-art performance with at least 2.7%/2.9% AUC and 2.6%/2.3% C-Index improvement with the patches labeled for two scales. Ablation study and visualization analysis further verify the effectiveness.

Xiao Pan, Zongxin Yang, Jianxin Ma, Chang Zhou, Yi Yang

In this paper, we focus on the task of generalizable neural human rendering which trains conditional Neural Radiance Fields (NeRF) from multi-view videos of different characters. To handle the dynamic human motion, previous methods have primarily used a SparseConvNet (SPC)-based human representation to process the painted SMPL. However, such SPC-based representation i) optimizes under the volatile observation space which leads to the pose-misalignment between training and inference stages, and ii) lacks the global relationships among human parts that is critical for handling the incomplete painted SMPL. Tackling these issues, we present a brand-new framework named TransHuman, which learns the painted SMPL under the canonical space and captures the global relationships between human parts with transformers. Specifically, TransHuman is mainly composed of Transformer-based Human Encoding (TransHE), Deformable Partial Radiance Fields (DPaRF), and Fine-grained Detail Integration (FDI). TransHE first processes the painted SMPL under the canonical space via transformers for capturing the global relationships between human parts. Then, DPaRF binds each output token with a deformable radiance field for encoding the query point under the observation space. Finally, the FDI is employed to further integrate fine-grained information from reference images. Extensive experiments on ZJU-MoCap and H36M show that our TransHuman achieves a significantly new state-of-the-art performance with high efficiency. Project page: https://pansanity666.github.io/TransHuman/

Ziyu Li, Jingming Guo, Tongtong Cao, Liu Bingbing, Wankou Yang

LiDAR-based 3D detection has made great progress in recent years. However, the performance of 3D detectors is considerably limited when deployed in unseen environments, owing to the severe domain gap problem. Existing domain adaptive 3D detection methods do not adequately consider the problem of the distributional discrepancy in feature space, thereby hindering the generalization of detectors across domains. In this work, we propose a novel unsupervised domain adaptive 3D detection framework, namely Geometry-aware Prototype Alignment (GPA-3D), which explicitly leverages the intrinsic geometric relationship from point cloud objects to reduce the feature discrepancy, thus facilitating cross-domain transferring. Specifically, GPA-3D assigns a series of tailored and learnable prototypes to point cloud objects with distinct geometric structures. Each prototype aligns BEV (bird's-eye-view) features derived from corresponding point cloud objects on source and target domains, reducing the distributional discrepancy and achieving better adaptation. The evaluation results obtained on various benchmarks, including Waymo, nuScenes and KITTI, demonstrate the superiority of our GPA-3D over the state-of-the-art approaches for different adaptation scenarios. The MindSpore version code will be publicly available at https://github.com/Liz66666/GPA3D.