Shuchang Zhou

Maxime Gasse, Quentin Cappart, Jonas Charfreitag et al. · deepmind, utoronto

Combinatorial optimization is a well-established area in operations research and computer science. Until recently, its methods have focused on solving problem instances in isolation, ignoring that they often stem from related data distributions in practice. However, recent years have seen a surge of interest in using machine learning as a new approach for solving combinatorial problems, either directly as solvers or by enhancing exact solvers. Based on this context, the ML4CO aims at improving state-of-the-art combinatorial optimization solvers by replacing key heuristic components. The competition featured three challenging tasks: finding the best feasible solution, producing the tightest optimality certificate, and giving an appropriate solver configuration. Three realistic datasets were considered: balanced item placement, workload apportionment, and maritime inventory routing. This last dataset was kept anonymous for the contestants.

Ruihang Miao, Weizhou Liu, Mingrui Chen et al.

3D Semantic Scene Completion (SSC) can provide dense geometric and semantic scene representations, which can be applied in the field of autonomous driving and robotic systems. It is challenging to estimate the complete geometry and semantics of a scene solely from visual images, and accurate depth information is crucial for restoring 3D geometry. In this paper, we propose the first stereo SSC method named OccDepth, which fully exploits implicit depth information from stereo images (or RGBD images) to help the recovery of 3D geometric structures. The Stereo Soft Feature Assignment (Stereo-SFA) module is proposed to better fuse 3D depth-aware features by implicitly learning the correlation between stereo images. In particular, when the input are RGBD image, a virtual stereo images can be generated through original RGB image and depth map. Besides, the Occupancy Aware Depth (OAD) module is used to obtain geometry-aware 3D features by knowledge distillation using pre-trained depth models. In addition, a reformed TartanAir benchmark, named SemanticTartanAir, is provided in this paper for further testing our OccDepth method on SSC task. Compared with the state-of-the-art RGB-inferred SSC method, extensive experiments on SemanticKITTI show that our OccDepth method achieves superior performance with improving +4.82% mIoU, of which +2.49% mIoU comes from stereo images and +2.33% mIoU comes from our proposed depth-aware method. Our code and trained models are available at https://github.com/megvii-research/OccDepth.

Shuangkang Fang, Weixin Xu, Heng Wang et al.

Neural Radiance Fields (NeRF) methods have proved effective as compact, high-quality and versatile representations for 3D scenes, and enable downstream tasks such as editing, retrieval, navigation, etc. Various neural architectures are vying for the core structure of NeRF, including the plain Multi-Layer Perceptron (MLP), sparse tensors, low-rank tensors, hashtables and their compositions. Each of these representations has its particular set of trade-offs. For example, the hashtable-based representations admit faster training and rendering but their lack of clear geometric meaning hampers downstream tasks like spatial-relation-aware editing. In this paper, we propose Progressive Volume Distillation (PVD), a systematic distillation method that allows any-to-any conversions between different architectures, including MLP, sparse or low-rank tensors, hashtables and their compositions. PVD consequently empowers downstream applications to optimally adapt the neural representations for the task at hand in a post hoc fashion. The conversions are fast, as distillation is progressively performed on different levels of volume representations, from shallower to deeper. We also employ special treatment of density to deal with its specific numerical instability problem. Empirical evidence is presented to validate our method on the NeRF-Synthetic, LLFF and TanksAndTemples datasets. For example, with PVD, an MLP-based NeRF model can be distilled from a hashtable-based Instant-NGP model at a 10X~20X faster speed than being trained the original NeRF from scratch, while achieving a superior level of synthesis quality. Code is available at https://github.com/megvii-research/AAAI2023-PVD.

Ailin Huang, Zhewei Huang, Shuchang Zhou

This paper reports our solution for ACM Multimedia ViCo 2022 Conversational Head Generation Challenge, which aims to generate vivid face-to-face conversation videos based on audio and reference images. Our solution focuses on training a generalized audio-to-head driver using regularization and assembling a high-visual quality renderer. We carefully tweak the audio-to-behavior model and post-process the generated video using our foreground-background fusion module. We get first place in the listening head generation track and second place in the talking head generation track on the official leaderboard. Our code is available at https://github.com/megvii-research/MM2022-ViCoPerceptualHeadGeneration.

Yun-Hao Cao, Peiqin Sun, Yechang Huang et al.

With the success of self-supervised learning (SSL), it has become a mainstream paradigm to fine-tune from self-supervised pretrained models to boost the performance on downstream tasks. However, we find that current SSL models suffer severe accuracy drops when performing low-bit quantization, prohibiting their deployment in resource-constrained applications. In this paper, we propose a method called synergistic self-supervised and quantization learning (SSQL) to pretrain quantization-friendly self-supervised models facilitating downstream deployment. SSQL contrasts the features of the quantized and full precision models in a self-supervised fashion, where the bit-width for the quantized model is randomly selected in each step. SSQL not only significantly improves the accuracy when quantized to lower bit-widths, but also boosts the accuracy of full precision models in most cases. By only training once, SSQL can then benefit various downstream tasks at different bit-widths simultaneously. Moreover, the bit-width flexibility is achieved without additional storage overhead, requiring only one copy of weights during training and inference. We theoretically analyze the optimization process of SSQL, and conduct exhaustive experiments on various benchmarks to further demonstrate the effectiveness of our method. Our code is available at https://github.com/megvii-research/SSQL-ECCV2022.

Yun-Hao Cao, Peiqin Sun, Shuchang Zhou

We propose universally slimmable self-supervised learning (dubbed as US3L) to achieve better accuracy-efficiency trade-offs for deploying self-supervised models across different devices. We observe that direct adaptation of self-supervised learning (SSL) to universally slimmable networks misbehaves as the training process frequently collapses. We then discover that temporal consistent guidance is the key to the success of SSL for universally slimmable networks, and we propose three guidelines for the loss design to ensure this temporal consistency from a unified gradient perspective. Moreover, we propose dynamic sampling and group regularization strategies to simultaneously improve training efficiency and accuracy. Our US3L method has been empirically validated on both convolutional neural networks and vision transformers. With only once training and one copy of weights, our method outperforms various state-of-the-art methods (individually trained or not) on benchmarks including recognition, object detection and instance segmentation. Our code is available at https://github.com/megvii-research/US3L-CVPR2023.

Shuangkang Fang, Yufeng Wang, Yi Yang et al.

Neural Radiance Fields (NeRF) have been widely adopted as practical and versatile representations for 3D scenes, facilitating various downstream tasks. However, different architectures, including the plain Multi-Layer Perceptron (MLP), Tensors, low-rank Tensors, Hashtables, and their combinations, entail distinct trade-offs. For instance, representations based on Hashtables enable faster rendering but lack clear geometric meaning, thereby posing challenges for spatial-relation-aware editing. To address this limitation and maximize the potential of each architecture, we propose Progressive Volume Distillation with Active Learning (PVD-AL), a systematic distillation method that enables any-to-any conversion between diverse architectures. PVD-AL decomposes each structure into two parts and progressively performs distillation from shallower to deeper volume representation, leveraging effective information retrieved from the rendering process. Additionally, a three-level active learning technique provides continuous feedback from teacher to student during the distillation process, achieving high-performance outcomes. Experimental evidence showcases the effectiveness of our method across multiple benchmark datasets. For instance, PVD-AL can distill an MLP-based model from a Hashtables-based model at a 10~20X faster speed and 0.8dB~2dB higher PSNR than training the MLP-based model from scratch. Moreover, PVD-AL permits the fusion of diverse features among distinct structures, enabling models with multiple editing properties and providing a more efficient model to meet real-time requirements like mobile devices. Project website: https://sk-fun.fun/PVD-AL.

Shuangkang Fang, Yufeng Wang, Yi Yang et al.

Numerous diffusion models have recently been applied to image synthesis and editing. However, editing 3D scenes is still in its early stages. It poses various challenges, such as the requirement to design specific methods for different editing types, retraining new models for various 3D scenes, and the absence of convenient human interaction during editing. To tackle these issues, we introduce a text-driven editing method, termed DN2N, which allows for the direct acquisition of a NeRF model with universal editing capabilities, eliminating the requirement for retraining. Our method employs off-the-shelf text-based editing models of 2D images to modify the 3D scene images, followed by a filtering process to discard poorly edited images that disrupt 3D consistency. We then consider the remaining inconsistency as a problem of removing noise perturbation, which can be solved by generating training data with similar perturbation characteristics for training. We further propose cross-view regularization terms to help the generalized NeRF model mitigate these perturbations. Our text-driven method allows users to edit a 3D scene with their desired description, which is more friendly, intuitive, and practical than prior works. Empirical results show that our method achieves multiple editing types, including but not limited to appearance editing, weather transition, material changing, and style transfer. Most importantly, our method generalizes well with editing abilities shared among a set of model parameters without requiring a customized editing model for some specific scenes, thus inferring novel views with editing effects directly from user input. The project website is available at https://sk-fun.fun/DN2N

Xiaotao Hu, Zhewei Huang, Ailin Huang et al.

The performance of video prediction has been greatly boosted by advanced deep neural networks. However, most of the current methods suffer from large model sizes and require extra inputs, e.g., semantic/depth maps, for promising performance. For efficiency consideration, in this paper, we propose a Dynamic Multi-scale Voxel Flow Network (DMVFN) to achieve better video prediction performance at lower computational costs with only RGB images, than previous methods. The core of our DMVFN is a differentiable routing module that can effectively perceive the motion scales of video frames. Once trained, our DMVFN selects adaptive sub-networks for different inputs at the inference stage. Experiments on several benchmarks demonstrate that our DMVFN is an order of magnitude faster than Deep Voxel Flow and surpasses the state-of-the-art iterative-based OPT on generated image quality. Our code and demo are available at https://huxiaotaostasy.github.io/DMVFN/.

Shengchao Zhou, Weizhou Liu, Chen Hu et al.

In the field of 3D object detection for autonomous driving, the sensor portfolio including multi-modality and single-modality is diverse and complex. Since the multi-modal methods have system complexity while the accuracy of single-modal ones is relatively low, how to make a tradeoff between them is difficult. In this work, we propose a universal cross-modality knowledge distillation framework (UniDistill) to improve the performance of single-modality detectors. Specifically, during training, UniDistill projects the features of both the teacher and the student detector into Bird's-Eye-View (BEV), which is a friendly representation for different modalities. Then, three distillation losses are calculated to sparsely align the foreground features, helping the student learn from the teacher without introducing additional cost during inference. Taking advantage of the similar detection paradigm of different detectors in BEV, UniDistill easily supports LiDAR-to-camera, camera-to-LiDAR, fusion-to-LiDAR and fusion-to-camera distillation paths. Furthermore, the three distillation losses can filter the effect of misaligned background information and balance between objects of different sizes, improving the distillation effectiveness. Extensive experiments on nuScenes demonstrate that UniDistill effectively improves the mAP and NDS of student detectors by 2.0%~3.2%.

Miao Fan, Mingrui Chen, Chen Hu et al.

Image matching is a fundamental and critical task in various visual applications, such as Simultaneous Localization and Mapping (SLAM) and image retrieval, which require accurate pose estimation. However, most existing methods ignore the occlusion relations between objects caused by camera motion and scene structure. In this paper, we propose Occ$^2$Net, a novel image matching method that models occlusion relations using 3D occupancy and infers matching points in occluded regions. Thanks to the inductive bias encoded in the Occupancy Estimation (OE) module, it greatly simplifies bootstrapping of a multi-view consistent 3D representation that can then integrate information from multiple views. Together with an Occlusion-Aware (OA) module, it incorporates attention layers and rotation alignment to enable matching between occluded and visible points. We evaluate our method on both real-world and simulated datasets and demonstrate its superior performance over state-of-the-art methods on several metrics, especially in occlusion scenarios.

Ran Ran, Jiwei Wei, Shuchang Zhou et al.

Video Temporal Grounding (VTG) faces a cross-modal semantic gap that often leads to background features being incorrectly aligned with the query, while directly matching the query to moments results in insufficient discriminability and consistency of temporal semantics. To address this issue, we propose MLLM-Assisted Semantic-Relational Consistent Alignment (MASRA), a training-time MLLM-based optimization framework for VTG. MASRA leverages an MLLM during training to produce two forms of textual priors, namely event-level descriptions with temporal spans and clip-level captions, and instantiates two MLLM-assisted alignments. Event Semantic Temporal Alignment (ESTA) aligns temporal context with event semantics to explicitly strengthen the correspondence between semantics and temporal events and improve span-level separability. Local Relational Consistency Alignment (LRCA) constructs a textual relation matrix derived from clip-level captions and aligns it with the temporal feature similarity matrix in the model, enhancing temporal consistency while capturing local structural information. MASRA includes two simple supporting modules, semantic-guided enhancement and second-order relational attention, to better utilize the learned semantic context and relational structure. Moreover, we introduce Decoupled Alignment Interaction (DAI) with a context-aware codebook to adaptively absorb query-irrelevant semantics and alleviate the cross-modal gap. The MLLM is only invoked during training and is not used at inference. Extensive experiments show that MASRA outperforms existing methods, and ablation studies validate its effectiveness.

Zhewei Huang, Ailin Huang, Xiaotao Hu et al.

The Space-Time Video Super-Resolution (STVSR) task aims to enhance the visual quality of videos, by simultaneously performing video frame interpolation (VFI) and video super-resolution (VSR). However, facing the challenge of the additional temporal dimension and scale inconsistency, most existing STVSR methods are complex and inflexible in dynamically modeling different motion amplitudes. In this work, we find that choosing an appropriate processing scale achieves remarkable benefits in flow-based feature propagation. We propose a novel Scale-Adaptive Feature Aggregation (SAFA) network that adaptively selects sub-networks with different processing scales for individual samples. Experiments on four public STVSR benchmarks demonstrate that SAFA achieves state-of-the-art performance. Our SAFA network outperforms recent state-of-the-art methods such as TMNet and VideoINR by an average improvement of over 0.5dB on PSNR, while requiring less than half the number of parameters and only 1/3 computational costs.

Shuchang Zhou, Shangkun Wu, Jiwei Wei et al.

AI-generated images are becoming increasingly realistic and diverse, posing significant challenges for generalizable detection. While Vision Foundation Models (VFMs) provide rich semantic representations and frequency-based methods capture complementary artifact cues, existing approaches that combine these modalities still suffer from limited generalization, with notable performance degradation on unseen generative models. We attribute this limitation to two key factors: frequency shortcut bias toward easily distinguishable cues associated with specific generators and cross-domain representation conflict between high-level semantics and low-level frequency patterns. To address these issues, we propose a Frequency-aware Gated Injection Network (FGINet) to improve generalization. Specifically, we design a Band-Masked Frequency Encoder (BMFE) that applies cross-band masking in the frequency domain to reduce reliance on generator-specific patterns and encourage more diverse and generalizable representations. We further introduce a Layer-wise Gated Frequency Injection (LGFI) mechanism to progressively inject frequency cues into the VFM backbone with adaptive gating, aligning with its hierarchical abstraction and alleviating representation conflict. Moreover, we propose a Hyperspherical Compactness Learning (HCL) framework with a cosine margin objective to learn compact and well-separated representations. Extensive experiments demonstrate that FGINet achieves state-of-the-art performance and strong generalization across multiple challenging datasets.

Shuchang Zhou, Kaiwen Shen, Jiwei Wei et al.

The rapid evolution of generative models has enabled the creation of highly realistic and diverse synthetic images, posing significant challenges to reliable and generalizable Synthetic Image Detection (SID). However, existing detectors are typically trained on limited and biased datasets, resulting in poor generalization to unseen generators. To address this issue, we propose HiMix, a unified framework that enhances generalization by expanding the training distribution and promoting artifact-aware representations. Specifically, the Mixup-driven Distributional Augmentation (MDA) module constructs continuous transitional samples between real and fake images, improving coverage of low-confidence regions and exposing the model to more challenging samples, while the pixel-wise mixup operation smoothly perturbs semantics to enhance sensitivity to low-level artifacts. Moreover, the Hierarchical Artifact-aware Representation (HAR) module aggregates artifact information from both global and local levels through cross-layer integration and coarse-to-fine feature fusion, enabling the extraction of discriminative forgery representations under diverse distributions. Extensive experiments across multiple benchmarks demonstrate that HiMix achieves state-of-the-art performance, establishing well-separated logits for improved generalization to unseen forgeries.

Miao Fan, Chen Hu, Shuchang Zhou

The Reinforcement Learning from Human Feedback (RLHF) plays a pivotal role in shaping the impact of large language models (LLMs), contributing significantly to controlling output toxicity and selecting output styles, particularly as LLMs often harbor misleading content, highlighting the urgency to align them with human values for secure AI systems. The RLHF, characterized by complexity, instability, and sensitivity to hyperparameters, makes the evaluation of the reward model for complex tasks challenging, thereby further complicating the use of Proximal Policy Optimization (PPO). In this paper, we introduce a simple task designed to employ Gloden as a reward model that validates the effectiveness of PPO and inspires it, primarily explaining the task of utilizing PPO to manipulate the tokenizer length of the output generated by the model. Experiments confirm that PPO is not only effective in manipulating the output tokenizer length to a certain extent in this type of task but also exhibits facilitated training once the influence of the reward model effect is excluded, making it an exciting development.

Haoyang Huang, Guoqing Ma, Nan Duan et al.

We present Step-Video-TI2V, a state-of-the-art text-driven image-to-video generation model with 30B parameters, capable of generating videos up to 102 frames based on both text and image inputs. We build Step-Video-TI2V-Eval as a new benchmark for the text-driven image-to-video task and compare Step-Video-TI2V with open-source and commercial TI2V engines using this dataset. Experimental results demonstrate the state-of-the-art performance of Step-Video-TI2V in the image-to-video generation task. Both Step-Video-TI2V and Step-Video-TI2V-Eval are available at https://github.com/stepfun-ai/Step-Video-TI2V.

Fenghe Tang, Bingkun Nian, Jianrui Ding et al.

In clinical practice, medical image analysis often requires efficient execution on resource-constrained mobile devices. However, existing mobile models-primarily optimized for natural images-tend to perform poorly on medical tasks due to the significant information density gap between natural and medical domains. Combining computational efficiency with medical imaging-specific architectural advantages remains a challenge when developing lightweight, universal, and high-performing networks. To address this, we propose a mobile model called Mobile U-shaped Vision Transformer (Mobile U-ViT) tailored for medical image segmentation. Specifically, we employ the newly purposed ConvUtr as a hierarchical patch embedding, featuring a parameter-efficient large-kernel CNN with inverted bottleneck fusion. This design exhibits transformer-like representation learning capacity while being lighter and faster. To enable efficient local-global information exchange, we introduce a novel Large-kernel Local-Global-Local (LGL) block that effectively balances the low information density and high-level semantic discrepancy of medical images. Finally, we incorporate a shallow and lightweight transformer bottleneck for long-range modeling and employ a cascaded decoder with downsample skip connections for dense prediction. Despite its reduced computational demands, our medical-optimized architecture achieves state-of-the-art performance across eight public 2D and 3D datasets covering diverse imaging modalities, including zero-shot testing on four unseen datasets. These results establish it as an efficient yet powerful and generalization solution for mobile medical image analysis. Code is available at https://github.com/FengheTan9/Mobile-U-ViT.

Zixuan Cao, Yang Xu, Zhewei Huang et al.

The Machine Learning for Combinatorial Optimization (ML4CO) NeurIPS 2021 competition aims to improve state-of-the-art combinatorial optimization solvers by replacing key heuristic components with machine learning models. On the dual task, we design models to make branching decisions to promote the dual bound increase faster. We propose a knowledge inheritance method to generalize knowledge of different models from the dataset aggregation process, named KIDA. Our improvement overcomes some defects of the baseline graph-neural-networks-based methods. Further, we won the $1$\textsuperscript{st} Place on the dual task. We hope this report can provide useful experience for developers and researchers. The code is available at https://github.com/megvii-research/NeurIPS2021-ML4CO-KIDA.

Yang Lin, Tianyu Zhang, Peiqin Sun et al.

Network quantization significantly reduces model inference complexity and has been widely used in real-world deployments. However, most existing quantization methods have been developed mainly on Convolutional Neural Networks (CNNs), and suffer severe degradation when applied to fully quantized vision transformers. In this work, we demonstrate that many of these difficulties arise because of serious inter-channel variation in LayerNorm inputs, and present, Power-of-Two Factor (PTF), a systematic method to reduce the performance degradation and inference complexity of fully quantized vision transformers. In addition, observing an extreme non-uniform distribution in attention maps, we propose Log-Int-Softmax (LIS) to sustain that and simplify inference by using 4-bit quantization and the BitShift operator. Comprehensive experiments on various transformer-based architectures and benchmarks show that our Fully Quantized Vision Transformer (FQ-ViT) outperforms previous works while even using lower bit-width on attention maps. For instance, we reach 84.89% top-1 accuracy with ViT-L on ImageNet and 50.8 mAP with Cascade Mask R-CNN (Swin-S) on COCO. To our knowledge, we are the first to achieve lossless accuracy degradation (~1%) on fully quantized vision transformers. The code is available at https://github.com/megvii-research/FQ-ViT.

Weixin Xu, Zipeng Feng, Shuangkang Fang et al.

Vast requirement of computation power of Deep Neural Networks is a major hurdle to their real world applications. Many recent Application Specific Integrated Circuit (ASIC) chips feature dedicated hardware support for Neural Network Acceleration. However, as ASICs take multiple years to develop, they are inevitably out-paced by the latest development in Neural Architecture Research. For example, Transformer Networks do not have native support on many popular chips, and hence are difficult to deploy. In this paper, we propose Arch-Net, a family of Neural Networks made up of only operators efficiently supported across most architectures of ASICs. When a Arch-Net is produced, less common network constructs, like Layer Normalization and Embedding Layers, are eliminated in a progressive manner through label-free Blockwise Model Distillation, while performing sub-eight bit quantization at the same time to maximize performance. Empirical results on machine translation and image classification tasks confirm that we can transform latest developed Neural Architectures into fast running and as-accurate Arch-Net, ready for deployment on multiple mass-produced ASIC chips. The code will be available at https://github.com/megvii-research/Arch-Net.

Pengbo Liu, Li Xiao, S. Kevin Zhou

There exists a large number of datasets for organ segmentation, which are partially annotated, and sequentially constructed. A typical dataset is constructed at a certain time by curating medical images and annotating the organs of interest. In other words, new datasets with annotations of new organ categories are built over time. To unleash the potential behind these partially labeled, sequentially-constructed datasets, we propose to learn a multi-organ segmentation model through incremental learning (IL). In each IL stage, we lose access to the previous annotations, whose knowledge is assumingly captured by the current model, and gain the access to a new dataset with annotations of new organ categories, from which we learn to update the organ segmentation model to include the new organs. We give the first attempt to conjecture that the different distribution is the key reason for 'catastrophic forgetting' that commonly exists in IL methods, and verify that IL has the natural adaptability to medical image scenarios. Extensive experiments on five open-sourced datasets are conducted to prove the effectiveness of our method and the conjecture mentioned above.

Zhewei Huang, Tianyuan Zhang, Wen Heng et al.

Real-time video frame interpolation (VFI) is very useful in video processing, media players, and display devices. We propose RIFE, a Real-time Intermediate Flow Estimation algorithm for VFI. To realize a high-quality flow-based VFI method, RIFE uses a neural network named IFNet that can estimate the intermediate flows end-to-end with much faster speed. A privileged distillation scheme is designed for stable IFNet training and improve the overall performance. RIFE does not rely on pre-trained optical flow models and can support arbitrary-timestep frame interpolation with the temporal encoding input. Experiments demonstrate that RIFE achieves state-of-the-art performance on several public benchmarks. Compared with the popular SuperSlomo and DAIN methods, RIFE is 4--27 times faster and produces better results. Furthermore, RIFE can be extended to wider applications thanks to temporal encoding. The code is available at https://github.com/megvii-research/ECCV2022-RIFE.

Zhewei Huang, Wen Heng, Shuchang Zhou

We show how to teach machines to paint like human painters, who can use a small number of strokes to create fantastic paintings. By employing a neural renderer in model-based Deep Reinforcement Learning (DRL), our agents learn to determine the position and color of each stroke and make long-term plans to decompose texture-rich images into strokes. Experiments demonstrate that excellent visual effects can be achieved using hundreds of strokes. The training process does not require the experience of human painters or stroke tracking data. The code is available at https://github.com/hzwer/ICCV2019-LearningToPaint.

Shuchang Zhou, Taihong Xiao, Yi Yang et al.

Object Transfiguration replaces an object in an image with another object from a second image. For example it can perform tasks like "putting exactly those eyeglasses from image A on the nose of the person in image B". Usage of exemplar images allows more precise specification of desired modifications and improves the diversity of conditional image generation. However, previous methods that rely on feature space operations, require paired data and/or appearance models for training or disentangling objects from background. In this work, we propose a model that can learn object transfiguration from two unpaired sets of images: one set containing images that "have" that kind of object, and the other set being the opposite, with the mild constraint that the objects be located approximately at the same place. For example, the training data can be one set of reference face images that have eyeglasses, and another set of images that have not, both of which spatially aligned by face landmarks. Despite the weak 0/1 labels, our model can learn an "eyeglasses" subspace that contain multiple representatives of different types of glasses. Consequently, we can perform fine-grained control of generated images, like swapping the glasses in two images by swapping the projected components in the "eyeglasses" subspace, to create novel images of people wearing eyeglasses. Overall, our deterministic generative model learns disentangled attribute subspaces from weakly labeled data by adversarial training. Experiments on CelebA and Multi-PIE datasets validate the effectiveness of the proposed model on real world data, in generating images with specified eyeglasses, smiling, hair styles, and lighting conditions etc. The code is available online.

Bohan Li, Jiazhe Guo, Hongsi Liu et al.

Generating high-fidelity, controllable, and annotated training data is critical for autonomous driving. Existing methods typically generate a single data form directly from a coarse scene layout, which not only fails to output rich data forms required for diverse downstream tasks but also struggles to model the direct layout-to-data distribution. In this paper, we introduce UniScene, the first unified framework for generating three key data forms - semantic occupancy, video, and LiDAR - in driving scenes. UniScene employs a progressive generation process that decomposes the complex task of scene generation into two hierarchical steps: (a) first generating semantic occupancy from a customized scene layout as a meta scene representation rich in both semantic and geometric information, and then (b) conditioned on occupancy, generating video and LiDAR data, respectively, with two novel transfer strategies of Gaussian-based Joint Rendering and Prior-guided Sparse Modeling. This occupancy-centric approach reduces the generation burden, especially for intricate scenes, while providing detailed intermediate representations for the subsequent generation stages. Extensive experiments demonstrate that UniScene outperforms previous SOTAs in the occupancy, video, and LiDAR generation, which also indeed benefits downstream driving tasks. Project page: https://arlo0o.github.io/uniscene/

Wenxun Wang, Shuchang Zhou, Wenyu Sun et al.

Transformers have shown remarkable performance in both natural language processing (NLP) and computer vision (CV) tasks. However, their real-time inference speed and efficiency are limited due to the inefficiency in Softmax and Layer Normalization (LayerNorm). Previous works based on function approximation suffer from inefficient implementation as they place emphasis on computation while disregarding memory overhead concerns. Moreover, such methods rely on retraining to compensate for approximation error which can be costly and inconvenient. In this paper, we present SOLE, a hardware-software co-design for Softmax and LayerNorm which is composed of E2Softmax and AILayerNorm. E2Softmax utilizes log2 quantization of exponent function and log-based division to approximate Softmax while AILayerNorm adopts low-precision statistic calculation. Compared with state-of-the-art designs, we achieve both low-precision calculation and low bit-width storage on Softmax and LayerNorm. Experiments show that SOLE maintains inference accuracy without retraining while offering orders of magnitude speedup and energy savings over GPU, achieving 3.04x, 3.86x energy-efficiency improvements and 2.82x, 3.32x area-efficiency improvements over prior state-of-the-art custom hardware for Softmax and LayerNorm, respectively.

Xintao Wang, Heng Wang, Yifei Zhang et al.

Role-playing language agents (RPLAs) have emerged as promising applications of large language models (LLMs). However, simulating established characters presents a challenging task for RPLAs, due to the lack of authentic character datasets and nuanced evaluation methods using such data. In this paper, we present CoSER, a collection of a high-quality dataset, open models, and an evaluation protocol towards effective RPLAs of established characters. The CoSER dataset covers 17,966 characters from 771 renowned books. It provides authentic dialogues with real-world intricacies, as well as diverse data types such as conversation setups, character experiences and internal thoughts. Drawing from acting methodology, we introduce given-circumstance acting for training and evaluating role-playing LLMs, where LLMs sequentially portray multiple characters in book scenes. Using our dataset, we develop CoSER 8B and CoSER 70B, i.e., advanced open role-playing LLMs built on LLaMA-3.1 models. Extensive experiments demonstrate the value of the CoSER dataset for RPLA training, evaluation and retrieval. Moreover, CoSER 70B exhibits state-of-the-art performance surpassing or matching GPT-4o on our evaluation and three existing benchmarks, i.e., achieving 75.80% and 93.47% accuracy on the InCharacter and LifeChoice benchmarks respectively.

Ailin Huang, Bingxin Li, Bruce Wang et al.

Large Audio-Language Models (LALMs) have significantly advanced intelligent human-computer interaction, yet their reliance on text-based outputs limits their ability to generate natural speech responses directly, hindering seamless audio interactions. To address this, we introduce Step-Audio-AQAA, a fully end-to-end LALM designed for Audio Query-Audio Answer (AQAA) tasks. The model integrates a dual-codebook audio tokenizer for linguistic and semantic feature extraction, a 130-billion-parameter backbone LLM and a neural vocoder for high-fidelity speech synthesis. Our post-training approach employs interleaved token-output of text and audio to enhance semantic coherence and combines Direct Preference Optimization (DPO) with model merge to improve performance. Evaluations on the StepEval-Audio-360 benchmark demonstrate that Step-Audio-AQAA excels especially in speech control, outperforming the state-of-art LALMs in key areas. This work contributes a promising solution for end-to-end LALMs and highlights the critical role of token-based vocoder in enhancing overall performance for AQAA tasks.

Linfeng Xu, Fanman Meng, Qingbo Wu et al.

The application of activity recognition in the ``AI + Education" field is gaining increasing attention. However, current work mainly focuses on the recognition of activities in manually captured videos and a limited number of activity types, with little attention given to recognizing activities in surveillance images from real classrooms. Activity recognition in classroom surveillance images faces multiple challenges, such as class imbalance and high activity similarity. To address this gap, we constructed a novel multimodal dataset focused on classroom surveillance image activity recognition called ARIC (Activity Recognition In Classroom). The ARIC dataset has advantages of multiple perspectives, 32 activity categories, three modalities, and real-world classroom scenarios. In addition to the general activity recognition tasks, we also provide settings for continual learning and few-shot continual learning. We hope that the ARIC dataset can act as a facilitator for future analysis and research for open teaching scenarios. You can download preliminary data from https://ivipclab.github.io/publication_ARIC/ARIC.

Ruibo Ming, Jingwei Wu, Zhewei Huang et al.

Recent advancements in auto-regressive large language models (LLMs) have led to their application in video generation. This paper explores the use of Large Vision Models (LVMs) for video continuation, a task essential for building world models and predicting future frames. We introduce ARCON, a scheme that alternates between generating semantic and RGB tokens, allowing the LVM to explicitly learn high-level structural video information. We find high consistency in the RGB images and semantic maps generated without special design. Moreover, we employ an optical flow-based texture stitching method to enhance visual quality. Experiments in autonomous driving scenarios show that our model can consistently generate long videos.

Mingyue Zhao, Han Li, Li Fan et al.

Fully-supervised airway segmentation has accomplished significant triumphs over the years in aiding pre-operative diagnosis and intra-operative navigation. However, full voxel-level annotation constitutes a labor-intensive and time-consuming task, often plagued by issues such as missing branches, branch annotation discontinuity, or erroneous edge delineation. label-efficient solutions for airway extraction are rarely explored yet primarily demanding in medical practice. To this end, we introduce a novel skeleton-level annotation (SkA) tailored to the airway, which simplifies the annotation workflow while enhancing annotation consistency and accuracy, preserving the complete topology. Furthermore, we propose a skeleton-supervised learning framework to achieve accurate airway segmentation. Firstly, a dual-stream buffer inference is introduced to realize initial label propagation from SkA, avoiding the collapse of direct learning from SkA. Then, we construct a geometry-aware dual-path propagation framework (GDP) to further promote complementary propagation learning, composed of hard geometry-aware propagation learning and soft geometry-aware propagation guidance. Experiments reveal that our proposed framework outperforms the competing methods with SKA, which amounts to only 1.96% airways, and achieves comparable performance with the baseline model that is fully supervised with 100% airways, demonstrating its significant potential in achieving label-efficient segmentation for other tubular structures, such as vessels.

Yudi Sang, Yanzhen Liu, Sutuke Yibulayimu et al.

The segmentation of pelvic fracture fragments in CT and X-ray images is crucial for trauma diagnosis, surgical planning, and intraoperative guidance. However, accurately and efficiently delineating the bone fragments remains a significant challenge due to complex anatomy and imaging limitations. The PENGWIN challenge, organized as a MICCAI 2024 satellite event, aimed to advance automated fracture segmentation by benchmarking state-of-the-art algorithms on these complex tasks. A diverse dataset of 150 CT scans was collected from multiple clinical centers, and a large set of simulated X-ray images was generated using the DeepDRR method. Final submissions from 16 teams worldwide were evaluated under a rigorous multi-metric testing scheme. The top-performing CT algorithm achieved an average fragment-wise intersection over union (IoU) of 0.930, demonstrating satisfactory accuracy. However, in the X-ray task, the best algorithm attained an IoU of 0.774, highlighting the greater challenges posed by overlapping anatomical structures. Beyond the quantitative evaluation, the challenge revealed methodological diversity in algorithm design. Variations in instance representation, such as primary-secondary classification versus boundary-core separation, led to differing segmentation strategies. Despite promising results, the challenge also exposed inherent uncertainties in fragment definition, particularly in cases of incomplete fractures. These findings suggest that interactive segmentation approaches, integrating human decision-making with task-relevant information, may be essential for improving model reliability and clinical applicability.

Shuchang Zhou, Jiwei Wei, Shiyuan He et al.

Prompt tuning has become a popular strategy for adapting Vision-Language Models (VLMs) to zero/few-shot visual recognition tasks. Some prompting techniques introduce prior knowledge due to its richness, but when learnable tokens are randomly initialized and disconnected from prior knowledge, they tend to overfit on seen classes and struggle with domain shifts for unseen ones. To address this issue, we propose the InPK model, which infuses class-specific prior knowledge into the learnable tokens during initialization, thus enabling the model to explicitly focus on class-relevant information. Furthermore, to mitigate the weakening of class information by multi-layer encoders, we continuously reinforce the interaction between learnable tokens and prior knowledge across multiple feature levels. This progressive interaction allows the learnable tokens to better capture the fine-grained differences and universal visual concepts within prior knowledge, enabling the model to extract more discriminative and generalized text features. Even for unseen classes, the learned interaction allows the model to capture their common representations and infer their appropriate positions within the existing semantic structure. Moreover, we introduce a learnable text-to-vision projection layer to accommodate the text adjustments, ensuring better alignment of visual-text semantics. Extensive experiments on 11 recognition datasets show that InPK significantly outperforms state-of-the-art methods in multiple zero/few-shot image classification tasks.

Qiyuan Wang, Yanzhe Liu, Shang Zhao et al.

For robotic surgical videos, instrument presence annotations are typically recorded with video streams, which offering the potential to reduce the manually annotated costs for segmentation. However, weakly supervised surgical instrument segmentation with only instrument presence labels has been rarely explored in surgical domain due to the highly under-constrained challenges. Temporal properties can enhance representation learning by capturing sequential dependencies and patterns over time even in incomplete supervision situations. From this, we take the inherent temporal attributes of surgical video into account and extend a two-stage weakly supervised segmentation paradigm from different perspectives. Firstly, we make temporal equivariance constraint to enhance pixel-wise temporal consistency between adjacent features. Secondly, we constrain class-aware semantic continuity between global and local regions across temporal dimension. Finally, we generate temporal-enhanced pseudo masks from consecutive frames to suppress irrelevant regions. Extensive experiments are validated on two surgical video datasets, including one cholecystectomy surgery benchmark and one real robotic left lateral segment liver surgery dataset. We annotate instance-wise instrument labels with fixed time-steps which are double checked by a clinician with 3-years experience to evaluate segmentation results. Experimental results demonstrate the promising performances of our method, which consistently achieves comparable or favorable results with previous state-of-the-art approaches.

Ruibo Ming, Zhewei Huang, Jingwei Wu et al.

Future Frame Synthesis (FFS), the task of generating subsequent video frames from context, represents a core challenge in machine intelligence and a cornerstone for developing predictive world models. This survey provides a comprehensive analysis of the FFS landscape, charting its critical evolution from deterministic algorithms focused on pixel-level accuracy to modern generative paradigms that prioritize semantic coherence and dynamic plausibility. We introduce a novel taxonomy organized by algorithmic stochasticity, which not only categorizes existing methods but also reveals the fundamental drivers--advances in architectures, datasets, and computational scale--behind this paradigm shift. Critically, our analysis identifies a bifurcation in the field's trajectory: one path toward efficient, real-time prediction, and another toward large-scale, generative world simulation. By pinpointing key challenges and proposing concrete research questions for both frontiers, this survey serves as an essential guide for researchers aiming to advance the frontiers of visual dynamic modeling.

Yuanqi Du, Quan Quan, Hu Han et al.

Pseudo-normality synthesis, which computationally generates a pseudo-normal image from an abnormal one (e.g., with lesions), is critical in many perspectives, from lesion detection, data augmentation to clinical surgery suggestion. However, it is challenging to generate high-quality pseudo-normal images in the absence of the lesion information. Thus, expensive lesion segmentation data have been introduced to provide lesion information for the generative models and improve the quality of the synthetic images. In this paper, we aim to alleviate the need of a large amount of lesion segmentation data when generating pseudo-normal images. We propose a Semi-supervised Medical Image generative LEarning network (SMILE) which not only utilizes limited medical images with segmentation masks, but also leverages massive medical images without segmentation masks to generate realistic pseudo-normal images. Extensive experiments show that our model outperforms the best state-of-the-art model by up to 6% for data augmentation task and 3% in generating high-quality images. Moreover, the proposed semi-supervised learning achieves comparable medical image synthesis quality with supervised learning model, using only 50 of segmentation data.

Felix Denzinger, Michael Wels, Katharina Breininger et al.

Coronary CT angiography (CCTA) has established its role as a non-invasive modality for the diagnosis of coronary artery disease (CAD). The CAD-Reporting and Data System (CAD-RADS) has been developed to standardize communication and aid in decision making based on CCTA findings. The CAD-RADS score is determined by manual assessment of all coronary vessels and the grading of lesions within the coronary artery tree. We propose a bottom-up approach for fully-automated prediction of this score using deep-learning operating on a segment-wise representation of the coronary arteries. The method relies solely on a prior fully-automated centerline extraction and segment labeling and predicts the segment-wise stenosis degree and the overall calcification grade as auxiliary tasks in a multi-task learning setup. We evaluate our approach on a data collection consisting of 2,867 patients. On the task of identifying patients with a CAD-RADS score indicating the need for further invasive investigation our approach reaches an area under curve (AUC) of 0.923 and an AUC of 0.914 for determining whether the patient suffers from CAD. This level of performance enables our approach to be used in a fully-automated screening setup or to assist diagnostic CCTA reading, especially due to its neural architecture design -- which allows comprehensive predictions.

Dachao Lin, Peiqin Sun, Guangzeng Xie et al.

Quantized Neural Networks (QNNs) use low bit-width fixed-point numbers for representing weight parameters and activations, and are often used in real-world applications due to their saving of computation resources and reproducibility of results. Batch Normalization (BN) poses a challenge for QNNs for requiring floating points in reciprocal operations, and previous QNNs either require computing BN at high precision or revise BN to some variants in heuristic ways. In this work, we propose a novel method to quantize BN by converting an affine transformation of two floating points to a fixed-point operation with shared quantized scale, which is friendly for hardware acceleration and model deployment. We confirm that our method maintains same outputs through rigorous theoretical analysis and numerical analysis. Accuracy and efficiency of our quantization method are verified by experiments at layer level on CIFAR and ImageNet datasets. We also believe that our method is potentially useful in other problems involving quantization.

Zengming Shen, S. Kevin Zhou, Yifan Chen et al.

Recently image-to-image translation has attracted significant interests in the literature, starting from the successful use of the generative adversarial network (GAN), to the introduction of cyclic constraint, to extensions to multiple domains. However, in existing approaches, there is no guarantee that the mapping between two image domains is unique or one-to-one. Here we propose a self-inverse network learning approach for unpaired image-to-image translation. Building on top of CycleGAN, we learn a self-inverse function by simply augmenting the training samples by swapping inputs and outputs during training and with separated cycle consistency loss for each mapping direction. The outcome of such learning is a proven one-to-one mapping function. Our extensive experiments on a variety of datasets, including cross-modal medical image synthesis, object transfiguration, and semantic labeling, consistently demonstrate clear improvement over the CycleGAN method both qualitatively and quantitatively. Especially our proposed method reaches the state-of-the-art result on the cityscapes benchmark dataset for the label to photo unpaired directional image translation.

Wen Heng, Shuchang Zhou, Tingting Jiang

Adversarial attacks find perturbations that can fool models into misclassifying images. Previous works had successes in generating noisy/edge-rich adversarial perturbations, at the cost of degradation of image quality. Such perturbations, even when they are small in scale, are usually easily spottable by human vision. In contrast, we propose Harmonic Adversar- ial Attack Methods (HAAM), that generates edge-free perturbations by using harmonic functions. The property of edge-free guarantees that the generated adversarial images can still preserve visual quality, even when perturbations are of large magnitudes. Experiments also show that adversaries generated by HAAM often have higher rates of success when transferring between models. In addition, we find harmonic perturbations can simulate natural phenomena like natural lighting and shadows. It would then be possible to help find corner cases for given models, as a first step to improving them.

Zhewei Huang, Wen Heng, Yuanzheng Tao et al.

Background elimination for noisy character images or character images from real scene is still a challenging problem, due to the bewildering backgrounds, uneven illumination, low resolution and different distortions. We propose a stroke-based character reconstruction(SCR) method that use a weighted quadratic Bezier curve(WQBC) to represent strokes of a character. Only training on our synthetic data, our stroke extractor can achieve excellent reconstruction effect in real scenes. Meanwhile. It can also help achieve great ability in defending adversarial attacks of character recognizers.

Guangzeng Xie, Yitan Wang, Shuchang Zhou et al.

In this paper we explore acceleration techniques for large scale nonconvex optimization problems with special focuses on deep neural networks. The extrapolation scheme is a classical approach for accelerating stochastic gradient descent for convex optimization, but it does not work well for nonconvex optimization typically. Alternatively, we propose an interpolation scheme to accelerate nonconvex optimization and call the method Interpolatron. We explain motivation behind Interpolatron and conduct a thorough empirical analysis. Empirical results on DNNs of great depths (e.g., 98-layer ResNet and 200-layer ResNet) on CIFAR-10 and ImageNet show that Interpolatron can converge much faster than the state-of-the-art methods such as the SGD with momentum and Adam. Furthermore, Anderson's acceleration, in which mixing coefficients are computed by least-squares estimation, can also be used to improve the performance. Both Interpolatron and Anderson's acceleration are easy to implement and tune. We also show that Interpolatron has linear convergence rate under certain regularity assumptions.

Łukasz Kidziński, Sharada Prasanna Mohanty, Carmichael Ong et al.

In the NIPS 2017 Learning to Run challenge, participants were tasked with building a controller for a musculoskeletal model to make it run as fast as possible through an obstacle course. Top participants were invited to describe their algorithms. In this work, we present eight solutions that used deep reinforcement learning approaches, based on algorithms such as Deep Deterministic Policy Gradient, Proximal Policy Optimization, and Trust Region Policy Optimization. Many solutions use similar relaxations and heuristics, such as reward shaping, frame skipping, discretization of the action space, symmetry, and policy blending. However, each of the eight teams implemented different modifications of the known algorithms.

Shuchang Zhou, Yuzhi Wang, He Wen et al.

Quantized Neural Networks (QNNs), which use low bitwidth numbers for representing parameters and performing computations, have been proposed to reduce the computation complexity, storage size and memory usage. In QNNs, parameters and activations are uniformly quantized, such that the multiplications and additions can be accelerated by bitwise operations. However, distributions of parameters in Neural Networks are often imbalanced, such that the uniform quantization determined from extremal values may under utilize available bitwidth. In this paper, we propose a novel quantization method that can ensure the balance of distributions of quantized values. Our method first recursively partitions the parameters by percentiles into balanced bins, and then applies uniform quantization. We also introduce computationally cheaper approximations of percentiles to reduce the computation overhead introduced. Overall, our method improves the prediction accuracies of QNNs without introducing extra computation during inference, has negligible impact on training speed, and is applicable to both Convolutional Neural Networks and Recurrent Neural Networks. Experiments on standard datasets including ImageNet and Penn Treebank confirm the effectiveness of our method. On ImageNet, the top-5 error rate of our 4-bit quantized GoogLeNet model is 12.7\%, which is superior to the state-of-the-arts of QNNs.

Xinyu Zhou, Cong Yao, He Wen et al.

Previous approaches for scene text detection have already achieved promising performances across various benchmarks. However, they usually fall short when dealing with challenging scenarios, even when equipped with deep neural network models, because the overall performance is determined by the interplay of multiple stages and components in the pipelines. In this work, we propose a simple yet powerful pipeline that yields fast and accurate text detection in natural scenes. The pipeline directly predicts words or text lines of arbitrary orientations and quadrilateral shapes in full images, eliminating unnecessary intermediate steps (e.g., candidate aggregation and word partitioning), with a single neural network. The simplicity of our pipeline allows concentrating efforts on designing loss functions and neural network architecture. Experiments on standard datasets including ICDAR 2015, COCO-Text and MSRA-TD500 demonstrate that the proposed algorithm significantly outperforms state-of-the-art methods in terms of both accuracy and efficiency. On the ICDAR 2015 dataset, the proposed algorithm achieves an F-score of 0.7820 at 13.2fps at 720p resolution.

He Wen, Shuchang Zhou, Zhe Liang et al.

Fully convolutional neural networks give accurate, per-pixel prediction for input images and have applications like semantic segmentation. However, a typical FCN usually requires lots of floating point computation and large run-time memory, which effectively limits its usability. We propose a method to train Bit Fully Convolution Network (BFCN), a fully convolutional neural network that has low bit-width weights and activations. Because most of its computation-intensive convolutions are accomplished between low bit-width numbers, a BFCN can be accelerated by an efficient bit-convolution implementation. On CPU, the dot product operation between two bit vectors can be reduced to bitwise operations and popcounts, which can offer much higher throughput than 32-bit multiplications and additions. To validate the effectiveness of BFCN, we conduct experiments on the PASCAL VOC 2012 semantic segmentation task and Cityscapes. Our BFCN with 1-bit weights and 2-bit activations, which runs 7.8x faster on CPU or requires less than 1\% resources on FPGA, can achieve comparable performance as the 32-bit counterpart.

Qinyao He, He Wen, Shuchang Zhou et al.

Reducing bit-widths of weights, activations, and gradients of a Neural Network can shrink its storage size and memory usage, and also allow for faster training and inference by exploiting bitwise operations. However, previous attempts for quantization of RNNs show considerable performance degradation when using low bit-width weights and activations. In this paper, we propose methods to quantize the structure of gates and interlinks in LSTM and GRU cells. In addition, we propose balanced quantization methods for weights to further reduce performance degradation. Experiments on PTB and IMDB datasets confirm effectiveness of our methods as performances of our models match or surpass the previous state-of-the-art of quantized RNN.

Cong Yao, Xiang Bai, Nong Sang et al.

Recently, scene text detection has become an active research topic in computer vision and document analysis, because of its great importance and significant challenge. However, vast majority of the existing methods detect text within local regions, typically through extracting character, word or line level candidates followed by candidate aggregation and false positive elimination, which potentially exclude the effect of wide-scope and long-range contextual cues in the scene. To take full advantage of the rich information available in the whole natural image, we propose to localize text in a holistic manner, by casting scene text detection as a semantic segmentation problem. The proposed algorithm directly runs on full images and produces global, pixel-wise prediction maps, in which detections are subsequently formed. To better make use of the properties of text, three types of information regarding text region, individual characters and their relationship are estimated, with a single Fully Convolutional Network (FCN) model. With such predictions of text properties, the proposed algorithm can simultaneously handle horizontal, multi-oriented and curved text in real-world natural images. The experiments on standard benchmarks, including ICDAR 2013, ICDAR 2015 and MSRA-TD500, demonstrate that the proposed algorithm substantially outperforms previous state-of-the-art approaches. Moreover, we report the first baseline result on the recently-released, large-scale dataset COCO-Text.

Shuchang Zhou, Yuxin Wu, Zekun Ni et al.

We propose DoReFa-Net, a method to train convolutional neural networks that have low bitwidth weights and activations using low bitwidth parameter gradients. In particular, during backward pass, parameter gradients are stochastically quantized to low bitwidth numbers before being propagated to convolutional layers. As convolutions during forward/backward passes can now operate on low bitwidth weights and activations/gradients respectively, DoReFa-Net can use bit convolution kernels to accelerate both training and inference. Moreover, as bit convolutions can be efficiently implemented on CPU, FPGA, ASIC and GPU, DoReFa-Net opens the way to accelerate training of low bitwidth neural network on these hardware. Our experiments on SVHN and ImageNet datasets prove that DoReFa-Net can achieve comparable prediction accuracy as 32-bit counterparts. For example, a DoReFa-Net derived from AlexNet that has 1-bit weights, 2-bit activations, can be trained from scratch using 6-bit gradients to get 46.1\% top-1 accuracy on ImageNet validation set. The DoReFa-Net AlexNet model is released publicly.