Neural network binarization accelerates deep models by quantizing their weights and activations into 1-bit. However, there is still a huge performance gap between Binary Neural Networks (BNNs) and their full-precision (FP) counterparts. As the quantization error caused by weights binarization has been reduced in earlier works, the activations binarization becomes the major obstacle for further improvement of the accuracy. BNN characterises a unique and interesting structure, where the binary and latent FP activations exist in the same forward pass (i.e. Binarize(a_F) = a_B). To mitigate the information degradation caused by the binarization operation from FP to binary activations, we establish a novel contrastive learning framework while training BNNs through the lens of Mutual Information (MI) maximization. MI is introduced as the metric to measure the information shared between binary and the FP activations, which assists binarization with contrastive learning. Specifically, the representation ability of the BNNs is greatly strengthened via pulling the positive pairs with binary and full-precision activations from the same input samples, as well as pushing negative pairs from different samples (the number of negative pairs can be exponentially large). This benefits the downstream tasks, not only classification but also segmentation and depth estimation, etc. The experimental results show that our method can be implemented as a pile-up module on existing state-of-the-art binarization methods and can remarkably improve the performance over them on CIFAR-10/100 and ImageNet, in addition to the great generalization ability on NYUD-v2.
论文检索
输入标题、作者或关键词，从 5,999 篇学术成果中精准定位
Self-attention based models such as vision transformers (ViTs) have emerged as a very competitive architecture alternative to convolutional neural networks (CNNs) in computer vision. Despite increasingly stronger variants with ever-higher recognition accuracies, due to the quadratic complexity of self-attention, existing ViTs are typically demanding in computation and model size. Although several successful design choices (e.g., the convolutions and hierarchical multi-stage structure) of prior CNNs have been reintroduced into recent ViTs, they are still not sufficient to meet the limited resource requirements of mobile devices. This motivates a very recent attempt to develop light ViTs based on the state-of-the-art MobileNet-v2, but still leaves a performance gap behind. In this work, pushing further along this under-studied direction we introduce EdgeViTs, a new family of light-weight ViTs that, for the first time, enable attention-based vision models to compete with the best light-weight CNNs in the tradeoff between accuracy and on-device efficiency. This is realized by introducing a highly cost-effective local-global-local (LGL) information exchange bottleneck based on optimal integration of self-attention and convolutions. For device-dedicated evaluation, rather than relying on inaccurate proxies like the number of FLOPs or parameters, we adopt a practical approach of focusing directly on on-device latency and, for the first time, energy efficiency. Extensive experiments on image classification, object detection, and semantic segmentation validate the high efficiency of our EdgeViTs when compared to the state-of-the-art efficient CNNs and ViTs in terms of accuracy-efficiency tradeoff on mobile hardware. Specifically, we show that our models are Pareto-optimal when both accuracy-latency and accuracy-energy tradeoffs are considered, achieving strict dominance over other ViTs in almost all cases and competing with the most efficient CNNs.
We present StreamDEQ, a method that infers frame-wise representations on videos with minimal per-frame computation. In contrast to conventional methods where compute time grows at least linearly with the network depth, we aim to update the representations in a continuous manner. For this purpose, we leverage the recently emerging implicit layer model which infers the representation of an image by solving a fixed-point problem. Our main insight is to leverage the slowly changing nature of videos and use the previous frame representation as an initial condition on each frame. This scheme effectively recycles the recent inference computations and greatly reduces the needed processing time. Through extensive experimental analysis, we show that StreamDEQ is able to recover near-optimal representations in a few frames time, and maintain an up-to-date representation throughout the video duration. Our experiments on video semantic segmentation and video object detection show that StreamDEQ achieves on par accuracy with the baseline (standard MDEQ) while being more than 3x faster.
Knowledge distillation has been applied to various tasks successfully. The current distillation algorithm usually improves students’ performance by imitating the output of the teacher. This paper shows that teachers can also improve students’ representation power by guiding students’ feature recovery. From this point of view, we propose Masked Generative Distillation (MGD), which is simple: we mask random pixels of the student’s feature and force it to generate the teacher’s full feature through a simple block. MGD is a truly general feature-based distillation method, which can be utilized on various tasks, including image classification, object detection, semantic segmentation and instance segmentation. We experiment on different models with extensive datasets and the results show that all the students achieve excellent improvements. Notably, we boost ResNet-18 from 69.90% to 71.69% ImageNet top-1 accuracy, RetinaNet with ResNet-50 backbone from 37.4 to 41.0 Boundingbox mAP, SOLO based on ResNet-50 from 33.1 to 36.2 Mask mAP and DeepLabV3 based on ResNet-18 from 73.20 to 76.02 mIoU. Our codes are available at https://github.com/yzd-v/MGD.
This work presents a simple vision transformer design as a strong baseline for object localization and instance segmentation tasks. Transformers recently demonstrate competitive performance in image classification tasks. To adopt ViT to object detection and dense prediction tasks, many works inherit the multistage design from convolutional networks and highly customized ViT architectures. Behind this design, the goal is to pursue a better trade-off between computational cost and effective aggregation of multiscale global contexts. However, existing works adopt multistage architectural design as a black-box solution without a clear understanding of its true benefits. In this paper, we comprehensively study three architecture design choices on ViT -- spatial reduction, doubled channels, and multiscale features -- and demonstrate that a vanilla ViT architecture can fulfill this goal without handcrafting multiscale features, maintaining the original ViT design philosophy. We further complete a scaling rule to optimize our model’s trade-off on accuracy and computation cost / model size. By leveraging a constant feature resolution and hidden size throughout the encoder blocks, we propose a simple and compact ViT architecture called Universal Vision Transformer (UViT) that achieves strong performance on COCO object detection and instance segmentation benchmark. Our code will be released upon acceptance.
One fundamental challenge in building an instance segmentation model for a large number of classes in complex scenes is the lack of training examples, especially for rare objects. In this paper, we explore the possibility to increase the training examples without laborious data collection and annotation. We find that an abundance of instance segments can potentially be obtained freely from object-centric images, according to two insights: (i) an object-centric image usually contains one salient object in a simple background; (ii) objects from the same class often share similar appearances or similar contrasts to the background. Motivated by these insights, we propose a simple and scalable framework FreeSeg for extracting and leveraging these ""free"" object foreground segments to facilitate model training in long-tailed instance segmentation. Concretely, we investigate the similarity among object-centric images of the same class to propose candidate segments of foreground instances, followed by a novel ranking of segment quality. The resulting high-quality object segments can then be used to augment the existing long-tailed datasets, e.g., by copying and pasting the segments onto the original training images. Extensive experiments show that FreeSeg yields substantial improvements on top of strong baselines and achieves state-of-the-art accuracy for segmenting rare object categories. Our code is publicly available at https://github.com/czhang0528/FreeSeg.
Major advancements have been made in the field of object detection and segmentation recently. However, when it comes to rare categories, the state-of-the-art methods fail to detect them, resulting in a significant performance gap between rare and frequent categories. In this paper, we identify that Sigmoid or Softmax functions used in deep detectors are a major reason for low performance and are suboptimal for long-tailed detection and segmentation. To address this, we develop a Gumbel Optimized Loss (GOL), for long-tailed detection and segmentation. It aligns with the Gumbel distribution of rare classes in imbalanced datasets, considering the fact that most classes in long-tailed detection have low expected probability. The proposed GOL significantly outperforms the best state-of-the-art method by 1.1% on AP, and boosts the overall segmentation by 9.0% and detection by 8.0%, particularly improving detection of rare classes by 20.3%, compared to Mask-RCNN, on LVIS dataset. Code available at: https://github.com/kostas1515/ GOL.
Graph-based approaches have been becoming increasingly popular in road network extraction, in addition to segmentation-based methods. Road networks are represented as graph structures, being able to explicitly define the topology structures and avoid the ambiguity of segmentation masks, such as between a real junction area and multiple separate roads in different heights. In contrast to the bottom-up graph-based approaches, which rely on orientation information, we propose a novel top-down approach to generate road network graphs with a holistic model, namely TD-Road. We decompose road extraction as two subtasks: key point prediction and connectedness prediction. We directly apply graph structures (i.e., locations of node and connections between them) as training supervisions for neural networks and generate road graph outputs in inference, instead of learning some intermediate properties of a graph structure (e.g., orientations or distances for the next move). Our network integrates a relation inference module with key point prediction, to capture connections between neighboring points and outputs the final road graphs with no post-processing steps required. Extensive experiments are conducted on challenging datasets, including City-Scale and SpaceNet to show the effectiveness and simplicity of our method, that the proposed method achieves remarkable results compared with previous state-of-the-art methods.
We present a conceptually simple, flexible, and universal visual perception head for variant visual tasks, e.g., classification, object detection, instance segmentation and pose estimation, and different frameworks, such as one-stage or two-stage pipelines. Our approach effectively identifies an object in an image while simultaneously generating a high-quality bounding box or contour-based segmentation mask or set of keypoints. The method, called UniHead, views different visual perception tasks as the dispersible points learning via the transformer encoder architecture. Given a fixed spatial coordinate, UniHead adaptively scatters it to different spatial points and reasons about their relations by transformer encoder. It directly outputs the final set of predictions in the form of multiple points, allowing us to perform different visual tasks in different frameworks with the same head design. We show extensive evaluations on ImageNet classification and all three tracks of the COCO suite of challenges, including object detection, instance segmentation and pose estimation. Without bells and whistles, UniHead can unify these visual tasks via a single visual head design and achieve comparable performance compared to expert models developed for each task. We hope our simple and universal UniHead will serve as a solid baseline and help promote universal visual perception research. Code and models are available at https://github.com/Sense-X/UniHead.
Poles and building edges are frequently observable objects on urban roads, conveying reliable hints for various computer vision tasks. To repetitively extract them as features and perform association between discrete LiDAR frames for registration, we propose the first learning-based feature segmentation and description model for 3D lines in LiDAR point cloud. To train our model without the time consuming and tedious data labeling process, we first generate synthetic primitives for the basic appearance of target lines, and build an iterative line auto-labeling process to gradually refine line labels on real LiDAR scans. Our segmentation model can extract lines under arbitrary scale perturbations, and we use shared EdgeConv encoder layers to train the two segmentation and descriptor heads jointly. Base on the model, we can build a highly-available global registration module for point cloud registration, in conditions without initial transformation hints. Experiments have demonstrated that our line-based registration method is highly competitive to state-of-the-art point-based approaches. Our code is available at https://github.com/zxrzju/SuperLine3D.git.
We aim to improve the performance of regressing hand keypoints and segmenting pixel-level hand masks under new imaging conditions (e.g., outdoors) when we only have labeled images taken under very different conditions (e.g., indoors). In the real world, it is important that the model trained for both tasks works under various imaging conditions. However, their variation covered by existing labeled hand datasets is limited. Thus, it is necessary to adapt the model trained on the labeled images (source) to unlabeled images (target) with unseen imaging conditions. While self-training domain adaptation methods (i.e., learning from the unlabeled target images in a self-supervised manner) have been developed for both tasks, their training may degrade performance when the predictions on the target images are noisy. To avoid this, it is crucial to assign a low importance (confidence) weight to the noisy predictions during self-training. In this paper, we propose to utilize the divergence of two predictions to estimate the confidence of the target image for both tasks. These predictions are given from two separate networks, and their divergence helps identify the noisy predictions. To integrate our proposed confidence estimation into self-training, we propose a teacher-student framework where the two networks (teachers) provide supervision to a network (student) for self-training, and the teachers are learned from the student by knowledge distillation. Our experiments show its superiority over state-of-the-art methods in adaptation settings with different lighting, grasping objects, backgrounds, and camera viewpoints. Our method improves by 4% the multi-task score on HO3D compared to the latest adversarial adaptation method. We also validate our method on Ego4D, egocentric videos with rapid changes in imaging conditions outdoors.
BEVFormer: Learning Bird’s-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers
PDF ↗3D visual perception tasks, including 3D detection and map segmentation based on multi-camera images, are essential for autonomous driving systems. In this work, we present a new framework termed BEVFormer, which learns unified BEV representations with spatiotemporal transformers to support multiple autonomous driving perception tasks. In a nutshell, BEVFormer exploits both spatial and temporal information by interacting with spatial and temporal space through predefined grid-shaped BEV queries. To aggregate spatial information, we design spatial cross-attention that each BEV query extracts the spatial features from the regions of interest across camera views. For temporal information, we propose temporal self-attention to recurrently fuse the history BEV information. Our approach achieves the new state-of-the-art 56.9% in terms of NDS metric on the nuScenes test set, which is 9.0 points higher than previous best arts and on par with the performance of LiDAR-based baselines. We further show that BEVFormer remarkably improves the accuracy of velocity estimation and recall of objects under low visibility conditions. The code will be released at https://github.com/zhiqi-li/BEVFormer.
Large-scale object detection and instance segmentation faces a severe data imbalance. The finer-grained object classes become, the less frequent they appear in our datasets. However at test-time, we expect a detector that performs well for all classes and not just the most frequent ones. In this paper, we provide a theoretical understanding of the long-trail detection problem. We show how the commonly used mean average precision evaluation metric on an unknown test-set is bound by a margin-based binary classification error on a long-tailed object-detection training set. We optimize margin-based binary classification error with a novel surrogate objective called Effective Class-Margin Loss (ECM). The ECM loss is simple, theoretically well-motivated, and outperforms other heuristic counterparts on LVIS v1 benchmark over a wide range of architecture and detectors. Code is available at https://github.com/janghyuncho/ECM-Loss.
Object pose estimation is a crucial task in computer vision and augmented reality. One of its key challenges is the difficulty of annotation of real training data and the lack of textured CAD models. Therefore, pipelines which do not require CAD models and which can be trained with few labeled images are desirable. We propose a weakly-supervised approach for object pose estimation from RGB-D data using training sets composed of very few labeled images with pose annotations along with weakly-labeled images with ground truth segmentation masks without pose labels. We achieve this by learning to annotate weakly-labeled training data through shape alignment while simultaneously training a pose prediction network. Point cloud alignment is performed using structure and rotation-invariant feature-based losses. We further learn an implicit shape representation, which allows the method to work without the known CAD model and also contributes to pose alignment and pose refinement during training on weakly labeled images. The experimental evaluation shows that our method achieves state-of-the-art results on LineMOD, Occlusion-LineMOD and TLess despite being trained using relative poses and on only a fraction of labeled data used by the other methods. We also achieve comparable results to state-of-the-art RGB-D based pose estimation approaches even when further reducing the amount of unlabeled training data. In addition, our method works even if relative camera poses are given instead of object pose annotations which are typically easier to obtain.
Do deep learning models for instance segmentation generalize to novel objects in a systematic way? For classification, such behavior has been questioned. In this study, we aim to understand if certain design decisions such as framework, architecture or pre-training contribute to the semantic understanding of instance segmentation. To answer this question, we consider a special case of robustness and compare pre-trained models on a challenging benchmark for object-centric out-of-distribution texture. We do not introduce another method in this work. Instead, we take a step back and evaluate a broad range of existing literature. This includes Cascade and Mask R-CNN, Swin Transformer, BMask, YOLACT(++), DETR, BCNet, SOTR and SOLOv2. We find that YOLACT++, SOTR and SOLOv2 are significantly more robust to out-of-distribution texture than other frameworks. In addition, we show that deeper and dynamic architectures improve robustness whereas training schedules, data augmentation and pre-training have only a minor impact. In summary we evaluate 68 models on 61 versions of MS COCO for a total of 4148 evaluations.
High-quality structured data with rich annotations are critical components in intelligent vehicle systems dealing with road scenes. However, data curation and annotation require intensive investments and yield low-diversity scenarios. The recently growing interest in synthetic data raises questions about the scope of improvement in such systems and the amount of manual work still required to produce high volumes and variations of simulated data. This work proposes a synthetic data generation pipeline that utilizes existing datasets, like nuScenes, to address the difficulties and domain-gaps present in simulated datasets. We show that using annotations and visual cues from existing datasets, we can facilitate automated multi-modal data generation, mimicking real scene properties with high-fidelity, along with mechanisms to diversify samples in a physically meaningful way. We demonstrate improvements in mIoU metrics by presenting qualitative and quantitative experiments with real and synthetic data for semantic segmentation on the Cityscapes and KITTI-STEP datasets. All relevant code and data is released on github.
We introduce a few-shot localization dataset originating from photographers who authentically were trying to learn about the visual content in the images they took. It includes over 8,000 segmentations of 100 categories in over 4,000 images that were taken by people with visual impairments. Compared to existing few-shot object detection and instance segmentation datasets, our dataset is the first to locate holes in objects (e.g., found in 12.4% of our segmentations), it shows objects that occupy a much larger range of sizes relative to the images, and text is over five times more common in our objects (e.g., found in 24.7% of our segmentations). Analysis of two modern few-shot localization algorithms demonstrates that they generalize poorly to our new dataset. The algorithms commonly struggle to locate objects with holes, very small and very large objects, and objects lacking text. To encourage a larger community to work on these unsolved challenges, we publicly share our annotated few-shot dataset at http://anonymous.
Most efforts on lesion segmentation from CT slices focus on one specific lesion type. However, universal and multi-category lesion segmentation is more important because the diagnoses of different body parts are usually correlated and carried out simultaneously. The existing universal lesion segmentation methods are weakly-supervised due to the lack of pixel-level annotation data. To bring this field into the fully-supervised era, we establish a large-scale universal lesion segmentation dataset, SegLesion. We also propose a baseline method for this task. Considering that it is easy to encode CT slices owing to the limited CT scenarios, we propose a Knowledge Embedding Module (KEM) to adapt the concept of dictionary learning for this task. Specifically, KEM first learns the knowledge encoding of CT slices and then embeds the learned knowledge encoding into the deep features of a CT slice to increase the distinguishability. With KEM incorporated, a Knowledge Embedding Network (KEN) is designed for universal lesion segmentation. To extensively compare KEN to previous segmentation methods, we build a large benchmark for SegLesion. KEN achieves state-of-the-art performance and can thus serve as a strong baseline for future research. Data and code will be released.
A key algorithm for understanding the world is material segmentation, which assigns a label (metal, glass, etc.) to each pixel. We find that a model trained on existing data underperforms in some settings and propose to address this with a large-scale dataset of 3.2 million dense segments on 44,560 indoor and outdoor images, which is 23x more segments than existing data. Our data covers a more diverse set of scenes, objects, viewpoints and materials, and contains a more fair distribution of skin types. We show that a model trained on our data outperforms a state-of-the-art model across datasets and viewpoints. We propose a large-scale scene parsing benchmark and baseline of 0.729 per-pixel accuracy, 0.585 mean class accuracy and 0.420 mean IoU across 46 materials.
Transparent objects are ubiquitous in household settings and pose distinct challenges for visual sensing and perception systems. The optical properties of transparent objects leaves conventional 3D sensors alone unreliable for object depth and pose estimation. These challenges are highlighted by the shortage of large-scale RGB-Depth datasets focusing on transparent objects in real-world settings. In this work, we contribute a large-scale real-world RGB-Depth transparent object dataset named ClearPose to serve as a benchmark dataset for segmentation, scene-level depth completion, and object-centric pose estimation tasks. The ClearPose dataset contains over 350K labeled real-world RGB-Depth frames and 4M instance annotations covering 63 household objects. The dataset includes object categories commonly used in daily life under various lighting and occluding conditions as well as challenging test scenarios such as cases of occlusion by opaque or translucent objects, non-planar orientations, presence of liquids, etc. We benchmark several state-of-the-art depth completion and object pose estimation deep neural networks on ClearPose.