In this paper, we propose a novel layer-adaptive weight-pruning approach for Deep Neural Networks (DNNs) that addresses the challenge of optimizing the output distortion minimization while adhering to a target pruning ratio constraint. Our approach takes into account the collective influence of all layers to design a layer-adaptive pruning scheme. We discover and utilize a very important additivity property of output distortion caused by pruning weights on multiple layers. This property enables us to formulate the pruning as a combinatorial optimization problem and efficiently solve it through dynamic programming. By decomposing the problem into sub-problems, we achieve linear time complexity, making our optimization algorithm fast and feasible to run on CPUs. Our extensive experiments demonstrate the superiority of our approach over existing methods on the ImageNet and CIFAR-10 datasets. On CIFAR-10, our method achieves remarkable improvements, outperforming others by up to 1.0% for ResNet-32, 0.5% for VGG-16, and 0.7% for DenseNet-121 in terms of top-1 accuracy. On ImageNet, we achieve up to 4.7% and 4.6% higher top-1 accuracy compared to other methods for VGG-16 and ResNet-50, respectively. These results highlight the effectiveness and practicality of our approach for enhancing DNN performance through layer-adaptive weight pruning. Code will be available on https://github.com/Akimoto-Cris/RD_VIT_PRUNE.
论文检索
输入标题、作者或关键词，从 2,893 篇学术成果中精准定位
We tackle the task of reconstructing hand-object interactions from short video clips. Given an input video, our approach casts 3D inference as a per-video optimization and recovers a neural 3D representation of the object shape, as well as the time-varying motion and hand articulation. While the input video naturally provides some multi-view cues to guide 3D inference, these are insufficient on their own due to occlusions and limited viewpoint variations. To obtain accurate 3D, we augment the multi-view signals with generic data-driven priors to guide reconstruction. Specifically, we learn a diffusion network to model the conditional distribution of (geometric) renderings of objects conditioned on hand configuration and category label, and leverage it as a prior to guide the novel-view renderings of the reconstructed scene. We empirically evaluate our approach on egocentric videos across 6 object categories, and observe significant improvements over prior single-view and multi-view methods. Finally, we demonstrate our system's ability to reconstruct arbitrary clips from YouTube, showing both 1st and 3rd person interactions.
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.
Image super-resolution (SR) has witnessed extensive neural network designs from CNN to transformer architectures. However, prevailing SR models suffer from prohibitive memory footprint and intensive computations, which limits further deployment on edge devices. This work investigates the potential of network pruning for super-resolution to take advantage of off-the-shelf network designs and reduce the underlying computational overhead. Two main challenges remain in applying pruning methods for SR. First, the widely-used filter pruning technique reflects limited granularity and restricted adaptability to diverse network structures. Second, existing pruning methods generally operate upon a pre-trained network for the sparse structure determination, hard to get rid of dense model training in the traditional SR paradigm. To address these challenges, we adopt unstructured pruning with sparse models directly trained from scratch. Specifically, we propose a novel Iterative Soft Shrinkage-Percentage (ISS-P) method by optimizing the sparse structure of a randomly initialized network at each iteration and tweaking unimportant weights with a small amount proportional to the magnitude scale on-the-fly. We observe that the proposed ISS-P can dynamically learn sparse structures adapting to the optimization process and preserve the sparse model's trainability by yielding a more regularized gradient throughput. Experiments on benchmark datasets demonstrate the effectiveness of the proposed ISS-P over diverse network architectures. Code is available at https://github.com/Jiamian-Wang/Iterative-Soft-Shrinkage-SR
Space-filling curves (SFCs) act as a linearization approach to map data in higher dimensional space to lower dimensional space, which is used comprehensively in computer vision, such as image/point cloud compression, hashing and etc. Currently, researchers formulate the problem of searching for an optimal SFC to the problem of finding a single Hamiltonian circuit on the image grid graph. Existing methods adopt graph neural networks (GNN) for SFC search. By modeling the pixel grid as a graph, they first adopt GNN to predict the edge weights and then generate a minimum spanning tree (MST) based on the predictions, which is further used to construct the SFC. However, GNN-based methods suffer from high computational costs and memory footprint usage. Besides, MST generation is un-differentiable, which is infeasible to optimize via gradient descent. To remedy these issues, we propose a GNN-based SFC-search framework with a tailored algorithm that largely reduces computational cost of GNN. Additionally, we propose a siamese network learning scheme to optimize DNN-based models in an end-to-end fashion. Extensive experiments show that our proposed method outperforms both DNN-based methods and traditional SFCs, e.g. Hilbert curve, by a large margin on various benchmarks.
Graph convolutional networks (GCNs) enable end-to-end learning on graph structured data. However, many works assume a given graph structure. When the input graph is noisy or unavailable, one approach is to construct or learn a latent graph structure. These methods typically fix the choice of node degree for the entire graph, which is suboptimal. Instead, we propose a novel end-to-end differentiable graph generator which builds graph topologies where each node selects both its neighborhood and its size. Our module can be readily integrated into existing pipelines involving graph convolution operations, replacing the predetermined or existing adjacency matrix with one that is learned, and optimized, as part of the general objective. As such it is applicable to any GCN. We integrate our module into trajectory prediction, point cloud classification and node classification pipelines resulting in improved accuracy over other structure-learning methods across a wide range of datasets and GCN backbones.
Unaligned 2D to 3D Translation with Conditional Vector-Quantized Code Diffusion using Transformers
PDF ↗Generating 3D images of complex objects conditionally from a few 2D views is a difficult synthesis problem, compounded by issues such as domain gap and geometric misalignment. For instance, a unified framework such as Generative Adversarial Networks cannot achieve this unless they explicitly define both a domain-invariant and geometric-invariant joint latent distribution, whereas Neural Radiance Fields are generally unable to handle both issues as they optimize at the pixel level. By contrast, we propose a simple and novel 2D to 3D synthesis approach based on conditional diffusion with vector-quantized codes. Operating in an information-rich code space enables high-resolution 3D synthesis via full-coverage attention across the views. Specifically, we generate the 3D codes (e.g. for CT images) conditional on previously generated 3D codes and the entire codebook of two 2D views (e.g. 2D X-rays). Qualitative and quantitative results demonstrate state-of-the-art performance over specialized methods across varied evaluation criteria, including fidelity metrics such as density, coverage, and distortion metrics for two complex volumetric imagery datasets from in real-world scenarios.
Deep neural networks have delivered remarkable performance and have been widely used in various visual tasks. However, their huge sizes cause significant inconvenience for transmission and storage. Many previous studies have explored model size compression. However, these studies often approach various lossy and lossless compression methods in isolation, leading to challenges in achieving high compression ratios efficiently. This work proposes a post-training model size compression method that combines lossy and lossless compression in a unified way. We first propose a unified parametric weight transformation, which ensures different lossy compression methods can be performed jointly in a post-training manner. Then, a dedicated differentiable counter is introduced to guide the optimization of lossy compression to arrive at a more suitable point for later lossless compression. Additionally, our method can easily control a desired global compression ratio and allocate adaptive ratios for different layers. Finally, our method can achieve a stable 10 times compression ratio without sacrificing accuracy and a 20 times compression ratio with minor accuracy loss in a short time. Our code is available at https://github.com/ModelTC/L2_Compression.
Current model quantization methods have shown their promising capability in reducing storage space and computation complexity. However, due to the diversity of quantization forms supported by different hardware, one limitation of existing solutions is that usually require repeated optimization for different scenarios. How to construct a model with flexible quantization forms has been less studied. In this paper, we explore a one-shot network quantization regime, named Elastic Quantization Neural Networks (EQ-Net), which aims to train a robust weight-sharing quantization supernet. First of all, we propose an elastic quantization space (including elastic bit-width, granularity, and symmetry) to adapt to various mainstream quantitative forms. Secondly, we propose the Weight Distribution Regularization Loss (WDR-Loss) and Group Progressive Guidance Loss (GPG-Loss) to bridge the inconsistency of the distribution for weights and output logits in the elastic quantization space gap. Lastly, we incorporate genetic algorithms and the proposed Conditional Quantization-Aware Accuracy Predictor (CQAP) as an estimator to quickly search mixed-precision quantized neural networks in supernet. Extensive experiments demonstrate that our EQ-Net is close to or even better than its static counterparts as well as state-of-the-art robust bit-width methods. Code can be available at https://github.com/xuke225/EQ-Net.git
Scene reconstructions are often incomplete due to occlusions and limited viewpoints. There have been efforts to use semantic information for scene completion. However, the completed shapes may be rough and imprecise since respective methods rely on 3D convolution and/or lack effective shape constraints. To overcome these limitations, we propose a semantic scene completion method based on deformable deep implicit templates (DDIT). Specifically, we complete each segmented instance in a scene by deforming a template with a latent code. Such a template is expressed by a deep implicit function in the canonical frame. It abstracts the shape prior of a category, and thus can provide constraints on the overall shape of an instance. Latent code controls the deformation of template to guarantee fine details of an instance. For code prediction, we design a neural network that leverages both intra- and inter-instance information. We also introduce an algorithm to transform instances between the world and canonical frames based on geometric constraints and a hierarchical tree. To further improve accuracy, we jointly optimize the latent code and transformation by enforcing the zero-valued isosurface constraint. In addition, we establish a new dataset to solve different problems of existing datasets. Experiments showed that our DDIT outperforms state-of-the-art approaches.
Face recognition systems are increasingly being used in different applications. In such systems, some features (also known as embeddings or templates) are extracted from each face image. Then, the extracted templates are stored in the system's database during the enrollment stage and are later used for recognition. In this paper, we focus on template inversion attacks against face recognition systems and introduce a novel method (dubbed GaFaR) to reconstruct 3D face from facial templates. To this end, we use a geometry-aware generator network based on generative neural radiance fields (GNeRF), and learn a mapping from facial templates to the intermediate latent space of the generator network. We train our network with a semi-supervised learning approach using real and synthetic images simultaneously. For the real training data, we use a Generative Adversarial Network (GAN) based framework to learn the distribution of the latent space. For the synthetic training data, where we have the true latent code, we directly train in the latent space of the generator network. In addition, during the inference stage, we also propose optimization on the camera parameters to generate face images to improve the success attack rate (up to 17.14% in our experiments). We evaluate the performance of our method in the whitebox and blackbox attacks against state-of-the-art face recognition models on the LFW and MOBIO datasets. To our knowledge, this paper is the first work on 3D face reconstruction from facial templates. The project page is available at: https://www.idiap.ch/paper/gafar
Accurate intrinsic calibration is essential for camera-based 3D perception, yet, it typically requires targets of well-known geometry. Here, we propose a camera self-calibration approach that infers camera intrinsics during application, from monocular videos in the wild. We propose to explicitly model projection functions and multi-view geometry, while leveraging the capabilities of deep neural networks for feature extraction and matching. To achieve this, we build upon recent research on integrating bundle adjustment into deep learning models, and introduce a self-calibrating bundle adjustment layer. The self-calibrating bundle adjustment layer optimizes camera intrinsics through classical Gauss-Newton steps and can be adapted to different camera models without re-training. As a specific realization, we implemented this layer within the deep visual SLAM system DROID-SLAM, and show that the resulting model, DroidCalib, yields state-of-the-art calibration accuracy across multiple public datasets. Our results suggest that the model generalizes to unseen environments and different camera models, including significant lens distortion. Thereby, the approach enables performing 3D perception tasks without prior knowledge about the camera. Code is available at https://github.com/boschresearch/droidcalib.
One-shot Neural architecture search (One-shot NAS) has been proposed as a time-efficient approach to obtain optimal subnet architectures and weights under different complexity cases by training only once. However, the subnet performance obtained by weight sharing is often inferior to the performance achieved by retraining. In this paper, we investigate the performance gap and attribute it to the use of uniform sampling, which is a common approach in supernet training. Uniform sampling concentrates training resources on subnets with intermediate computational resources, which are sampled with high probability. However, subnets with different complexity regions require different optimal training strategies for optimal performance. To address the problem of uniform sampling, we propose ShiftNAS, a method that can adjust the sampling probability based on the complexity of subnets. We achieve this by evaluating the performance variation of subnets with different complexity and designing an architecture generator that can accurately and efficiently provide subnets with the desired complexity. Both the sampling probability and the architecture generator can be trained end-to-end in a gradient-based manner. With ShiftNAS, we can directly obtain the optimal model architecture and parameters for a given computational complexity. We evaluate our approach on multiple visual network models, including convolutional neural networks (CNNs) and vision transformers (ViTs), and demonstrate that ShiftNAS is model-agnostic. Experimental results on ImageNet show that ShiftNAS can improve the performance of one-shot NAS without additional computational consumption. Source codes are available at GitHub.
Quantization scale and bit-width are the most important parameters when considering how to quantize a neural network. Prior work focuses on optimizing quantization scales in a global manner through gradient methods (gradient descent & Hessian analysis). Yet, when applying perturbations to quantization scales, we observe a very jagged, highly non-smooth test loss landscape. In fact, small perturbations in quantization scale can greatly affect accuracy, yielding a 0.5-0.8% accuracy boost in 4-bit quantized vision transformers (ViTs). In this regime, gradient methods break down, since they cannot reliably reach local minima. In our work, dubbed Evol-Q, we use evolutionary search to effectively traverse the non-smooth landscape. Additionally, we propose using an infoNCE loss, which not only helps combat overfitting on the small (1,000 images) calibration dataset but also makes traversing such a highly non-smooth surface easier. Evol-Q improves the top-1 accuracy of a fully quantized ViT-Base by 10.30%, 0.78%, and 0.15% for 3-bit, 4-bit, and 8-bit weight quantization levels. Extensive experiments on a variety of CNN and ViT architectures further demonstrate its robustness in extreme quantization scenarios.
Regression-based methods for 3D human pose estimation directly predict the 3D pose parameters from a 2D image using deep networks. While achieving state-of-the-art performance on standard benchmarks, their performance degrades under occlusion. In contrast, optimization-based methods fit a parametric body model to 2D features in an iterative manner. The localized reconstruction loss can potentially make them robust to occlusion, but they suffer from the 2D-3D ambiguity. Motivated by the recent success of generative models in rigid object pose estimation, we propose 3D-aware Neural Body Fitting (3DNBF) - an approximate analysis-by-synthesis approach to 3D human pose estimation with SOTA performance and occlusion robustness. In particular, we propose a generative model of deep features based on a volumetric human representation with Gaussian ellipsoidal kernels emitting 3D pose-dependent feature vectors. The neural features are trained with contrastive learning to become 3D-aware and hence to overcome the 2D-3D ambiguity. Experiments show that 3DNBF outperforms other approaches on both occluded and standard benchmarks.
Despite being highly performant, deep neural networks might base their decisions on features that spuriously correlate with the provided labels, thus hurting generalization. To mitigate this, 'model guidance' has recently gained popularity, i.e. the idea of regularizing the models' explanations to ensure that they are "right for the right reasons". While various techniques to achieve such model guidance have been proposed, experimental validation of these approaches has thus far been limited to relatively simple and / or synthetic datasets. To better understand the effectiveness of the various design choices that have been explored in the context of model guidance, in this work we conduct an in-depth evaluation across various loss functions, attribution methods, models, and 'guidance depths' on the PASCAL VOC 2007 and MS COCO 2014 datasets. As annotation costs for model guidance can limit its applicability, we also place a particular focus on efficiency. Specifically, we guide the models via bounding box annotations, which are much cheaper to obtain than the commonly used segmentation masks, and evaluate the robustness of model guidance under limited (e.g. with only 1% of annotated images) or overly coarse annotations. Further, we propose using the EPG score as an additional evaluation metric and loss function ('Energy loss'). We show that optimizing for the Energy loss leads to models that exhibit a distinct focus on object-specific features, despite only using bounding box annotations that also include background regions. Lastly, we show that such model guidance can improve generalization under distribution shifts. Code available at: https://github.com/sukrutrao/Model-Guidance
Neural Scene Flow Prior (NSFP) is of significant interest to the vision community due to its inherent robustness to out-of-distribution (OOD) effects and its ability to deal with dense lidar points. The approach utilizes a coordinate neural network to estimate scene flow at runtime, without any training. However, it is up to 100 times slower than current state-of-the-art learning methods. In other applications such as image, video, and radiance function reconstruction innovations in speeding up the runtime performance of coordinate networks have centered upon architectural changes. In this paper, we demonstrate that scene flow is different---with the dominant computational bottleneck stemming from the loss function itself (i.e., Chamfer distance). Further, we rediscover the distance transform (DT) as an efficient, correspondence-free loss function that dramatically speeds up the runtime optimization. Our fast neural scene flow (FNSF) approach reports for the first time real-time performance comparable to learning methods, without any training or OOD bias on two of the largest open autonomous driving (AV) lidar datasets Waymo Open [62] and Argoverse [8].
Neural Radiance Fields (NeRFs) have revolutionized 3D scene synthesis. Voxel grids are commonly employed to enhance training or rendering speed, but they entail additional storage requirements. The large model size and high computational and memory demands impede their progress on resource-constrained devices, e.g., Microcontroller Units (MCUs). Besides, there is currently no NeRF rendering framework available on MCU devices. In this paper, we propose a NeRF method named MCUNeRF for 3D scene synthesis on MCU devices. The proposed MCUNeRF compresses voxel grids via a hybrid quantization algorithm merging learned step-size quantization (LSQ) and optimized product quantization (OPQ). To further reduce the model storage, we also propose a codebook-sharing method that renders multiple objects with a single quantization codebook. Then we implement a NeRF-based rendering framework for MCU devices, which leverages a low-bit neural network computation framework, i.e. CMSIS-NN, to accelerate the rendering progress. Extensive experiments on four datasets such as Synthetic-NeRF demonstrate that our proposed method could compress model data by 20-40 times with comparable rendering quality, which enables NeRF-based scene rendering on MCU devices with only 1M SRAM.
While large deep neural networks excel at general video analytics tasks, the significant demand on computing capacity makes them infeasible for real-time inference on resource-constrained end cameras. In this paper, we propose an edge-assisted framework that continuously updates the lightweight model deployed on the end cameras to achieve accurate predictions in adverse environments. This framework consists of three modules, namely, a key frame extractor, a trigger controller, and a retraining manager. The low-cost key frame extractor obtains frames that can best represent the current environment. Those frames are then transmitted and buffered as the retraining data for model update at the edge server. Once the trigger controller detects a significant accuracy drop in the selected frames, the retraining manager outputs the optimal retraining configuration balancing the accuracy and time cost. We prototype our system on two end devices of different computing capacities with one edge server. The results demonstrate that our approach significantly improves accuracy across all tested adverse environment scenarios (up to 24%) and reduces more than 50% of the retraining time compared to existing benchmarks.
Deep hashing has been extensively explored for image retrieval due to fast computation and efficient storage. Since conventional deep hashing methods are not suitable for the common scenario in real life that data exhibits a long-tailed distribution, several long-tailed hashing methods have been proposed recently. However, existing long-tail hashing methods seek to utilize fixed class centroids and cannot fully develop the discriminative ability of hash codes for tail-class samples. Specifically, fixed class centroids cannot characterize authentic semantics of tail classes or provide effective semantic information for hash codes learning under the long-tailed setting. To this end, we propose a novel Dual Dynamic Proxy Hashing Network (DDPHN) with two sets of learnable dynamic proxies, i.e. hash proxies and feature proxies, to improve the discrimination of hash codes for tail-class samples. Compared with fixed class centroids, learnable proxies can be optimized constantly via the proxy learning loss and depict accurate class semantics despite the scarcity of tail-class samples. Apart from low-dimensional binary hash proxies, we introduce high-dimensional continuous feature proxies that can describe semantic relationships more precisely, contributing to hash codes learning as well. To further leverage semantic information carried by proxies, we build a hypergraph by exploring neighborhood relationships in the feature space and then introduce a hypergraph neural network to transfer knowledge from proxies to samples in the Hamming space. Extensive experiments show the superiority of our learnable dynamic proxies and demonstrate that our method outperforms numerous deep hashing models and recent state-of-the-art long-tailed hashing methods.