Transformers have demonstrated considerable success across various domains but are constrained by their significant computational and memory requirements. This poses challenges for deployment on resource-constrained devices. Quantization, as an effective model compression method, can significantly reduce the operational time of Transformers on edge devices. Notably, Transformers display more substantial outliers than convolutional neural networks, leading to uneven feature distribution among different channels and tokens. To address this issue, we propose an adaptive outlier correction quantization (AOCQ) method for Transformers, which significantly alleviates the adverse effects of these outliers. AOCQ adjusts the notable discrepancies in channels and tokens across three levels: operator level, framework level, and loss level. We introduce a new operator that equivalently balances the activations across different channels and insert an extra stage to optimize the activation quantization step on the framework level. Additionally, we transfer the imbalanced activations across tokens and channels to the optimization of model weights on the loss level. Based on the theoretical study, our method can reduce the quantization error. The effectiveness of the proposed method is verified on various benchmark models and tasks. Surprisingly, DeiT-Base with 8-bit post-training quantization (PTQ) can achieve 81.57% accuracy with a 0.28 percentage point drop while enjoying 4× faster runtime. Furthermore, the weights of Swin and DeiT on several tasks, including classification and object detection, can be post-quantized to ultra-low 4 bits, with a minimal accuracy loss of 2%, while requiring nearly 8× less memory.

Most existing knowledge distillation methods for semantic segmentation focus on extracting various sophisticated knowledge from raw features. However, such knowledge is usually manually designed and relies on prior knowledge as in traditional feature engineering. In this paper, we aim to propose a simple and effective feature distillation method using raw features. To this end, we revisit the pioneering work in feature distillation, FitNets, which simply minimizes the mean squared error (MSE) loss between the teacher and student features. Our experiments show that this naive method yields good results, even surpassing some well-designed methods in some cases. However, it requires carefully tuning the weight of distillation loss. By decomposing the loss function of FitNets into a magnitude difference term and an angular difference term, we find the weight of the angular difference term is affected by the magnitudes of the teacher features and the student features. We experimentally show that the angular difference term plays a crucial role in feature distillation and the magnitude of the features produced by different models may vary significantly. Therefore, it is hard to determine a suitable loss weight for various models. To avoid the weight of the angular distillation term being affected by the magnitude of the features, we propose Angular Distillation and explore distilling angular information along different feature dimensions for semantic segmentation. Extensive experiments show that our simple method exhibits great robustness to hyper-parameters and achieves state-of-the-art distillation performance for semantic segmentation.

Real-time learning on real-world data streams with temporal relations is essential for intelligent agents. However, current online Continual Learning (CL) benchmarks adopt the mini-batch setting and are composed of temporally unrelated and disjoint tasks as well as pre-set class boundaries. In this paper, we delve into a real-world CL scenario for fresh recognition where algorithms are required to recognize a huge variety of products to facilitate the checkout speed. Products mainly consists of packaged cereals, seasonal fruits, and vegetables from local farms or shipped from overseas. Since algorithms process instance streams consisting of sequential images, we name this real-world CL problem as Instance-Based Continual Learning (IBCL) . Different from the current online CL setting, algorithms are required to perform instant testing and learning upon each incoming instance. Moreover, IBCL has no task boundaries or class boundaries and allows the evolution and the forgetting of old samples within each class. To promote the researches on real CL challenges, we propose the first real-world CL dataset coined the Continual Fresh Recognition (CFR) dataset, which consists of fresh recognition data streams (766 K labelled images in total) collected from 30 supermarkets. Based on the CFR dataset, we extensively evaluate the performance of current online CL methods under various settings and find that current prominent online CL methods operate at high latency and demand significant memory consumption to cache old samples for replaying. Therefore, we make the first attempt to design an efficient and effective Instant Training-Free Learning (ITFL) framework for IBCL. ITFL consists of feature extractors trained in the metric learning manner and reformulates CL as a temporal classification problem among several most similar classes. Unlike current online CL methods that cache image samples (150 KB per image) and rely on training to learn new knowledge, our framework only caches features (2 KB per image) and is free of training in deployment. Extensive evaluations across three datasets demonstrate that our method achieves comparable recognition accuracy to current methods with lower latency and less resource consumption. Our codes and datasets will be publicly available at https://github.com/detectRecog/IBCL .

The present paper introduces sparsely supervised instance segmentation, with the datasets being fully annotated bounding boxes and sparsely annotated masks. A direct solution to this task is self-training, which is not fully explored for instance segmentation yet. In this paper, we propose MaskBooster for sparsely supervised instance segmentation (SpSIS) with comprehensive usage of pseudo masks. MaskBooster is featured with (1) dynamic and progressive pseudo masks from an online updating teacher model, (2) refining binary pseudo masks with the help of bounding box prior, (3) learning inter-class prediction distribution via knowledge distillation for soft pseudo masks. As an end-to-end and universal self-training framework, MaskBooster can empower fully supervised algorithms and boost their segmentation performance on SpSIS. Abundant experiments are conducted on COCO and BDD100K datasets and validate the effectiveness of MaskBooster. Specifically, on different COCO protocols and BDD100K, we surpass sparsely supervised baseline by a large margin for both Mask RCNN and ShapeProp. MaskBooster on SpSIS also outperforms weakly and semi-supervised instance segmentation state-of-the-art on the datasets with similar annotation budgets.

Post-training quantization (PTQ) is an effective compression method to reduce the model size and computational cost. However, quantizing a model into a low-bit one, e.g., lower than 4, is difficult and often results in non-negligible performance degradation. To address this, we investigate the loss landscapes of quantized networks with various bit-widths. We show that the network with more ragged loss surface, is more easily trapped into bad local minima, which mostly appears in low-bit quantization. A deeper analysis indicates, the ragged surface is caused by the injection of excessive quantization noise. To this end, we detach a sharpness term from the loss which reflects the impact of quantization noise. To smooth the rugged loss surface, we propose to limit the sharpness term small and stable during optimization. Instead of directly optimizing the target bit network, we design a self-adapted shrinking scheduler for the bit-width in continuous domain from high bit-width to the target by limiting the increasing sharpness term within a proper range. It can be viewed as iteratively adding small “instant” quantization noise and adjusting the network to eliminate its impact. Widely experiments including classification and detection tasks demonstrate the effectiveness of the Bit-shrinking strategy in PTQ. On the Vision Transformer models, our INT8 and INT6 models drop within 0.5% and 1.5% Top-1 accuracy, respectively. On the traditional CNN networks, our INT4 quantized models drop within 1.3% and 3.5% Top-1 accuracy on ResNet18 and MobileNetV2 without fine-tuning, which achieves the state-of-the-art performance.

Current methods for open-vocabulary object detection (OVOD) rely on a pre-trained vision-language model (VLM) to acquire the recognition ability. In this paper, we propose a simple yet effective framework to Distill the Knowledge from the VLM to a DETR-like detector, termed DK-DETR. Specifically, we present two ingenious distillation schemes named semantic knowledge distillation (SKD) and relational knowledge distillation (RKD). To utilize the rich knowledge from the VLM systematically, SKD transfers the semantic knowledge explicitly, while RKD exploits implicit relationship information between objects. Furthermore, a distillation branch including a group of auxiliary queries is added to the detector to mitigate the negative effect on base categories. Equipped with SKD and RKD on the distillation branch, DK-DETR improves the detection performance of novel categories significantly and avoids disturbing the detection of base categories. Extensive experiments on LVIS and COCO datasets show that DK-DETR surpasses existing OVOD methods under the setting that the base-category supervision is solely available. The code and models are available at https://github.com/hikvision-research/opera.

Current 6D pose estimation methods focus on handling objects that are previously trained, which limits their applications in real dynamic world. To this end, we propose a geometry correspondence-based framework, termed GCPose, to estimate 6D pose of arbitrary unseen objects without any re-training. Specifically, the proposed method draws the idea from point cloud registration and resorts to object-agnostic geometry features to establish the 3D-3D correspondences between the object-scene point cloud and object-model point cloud. Then the 6D pose parameters are solved by a least-squares fitting algorithm. Taking the symmetry properties of objects into consideration, we design a symmetry-aware matching loss to facilitate the learning of dense point-wise geometry features and improve the performance considerably. Moreover, we introduce an online training data generation with special data augmentation and normalization to empower the network to learn diverse geometry prior. With training on synthetic objects from ShapeNet, our method outperforms previous approaches for unseen object pose estimation by a large margin on T-LESS, LINEMOD, Occluded-LINEMOD, and TUD-L datasets. Code is available at https://github.com/hikvision-research/GCPose.

This paper presents an end-to-end instance segmentation framework, termed SOIT, that Segments Objects with Instance-aware Transformers. Inspired by DETR, our method views instance segmentation as a direct set prediction problem and effectively removes the need for many hand-crafted components like RoI cropping, one-to-many label assignment, and non-maximum suppression (NMS). In SOIT, multiple queries are learned to directly reason a set of object embeddings of semantic category, bounding-box location, and pixel-wise mask in parallel under the global image context. The class and bounding-box can be easily embedded by a fixed-length vector. The pixel-wise mask, especially, is embedded by a group of parameters to construct a lightweight instance-aware transformer. Afterward, a full-resolution mask is produced by the instance-aware transformer without involving any RoI-based operation. Overall, SOIT introduces a simple single-stage instance segmentation framework that is both RoI- and NMS-free. Experimental results on the MS COCO dataset demonstrate that SOIT outperforms state-of-the-art instance segmentation approaches significantly. Moreover, the joint learning of multiple tasks in a unified query embedding can also substantially improve the detection performance. Code is available at https://github.com/yuxiaodongHRI/SOIT.

Pseudo bounding boxes from the self-training paradigm are inevitably noisy for semi-supervised object detection. To cope with that, a dual decoupling training framework is proposed in the present study, i.e. clean and noisy data decoupling, and classification and localization task decoupling. In the first decoupling, two-level thresholds are used to categorize pseudo boxes into three groups, i.e. clean backgrounds, noisy foregrounds and clean foregrounds. With a specially designed noise-bypass head focusing on noisy data, backbone networks can extract coarse but diverse information; and meanwhile, an original head learns from clean samples for more precise predictions. In the second decoupling, we take advantage of the two-head structure for better evaluation of localization quality, thus the category label and location of a pseudo box can remain independent of each other during training. The approach of two-level thresholds is also applied to group pseudo boxes into three sections of different location accuracy. We outperform existing works by a large margin on VOC datasets, reaching 54.8 mAP(+1.8), and even up to 55.9 mAP(+1.5) by leveraging MS-COCO train2017 as extra unlabeled data. On MS-COCO benchmark, our method also achieves about 1.0 mAP improvements averaging across protocols compared with the prior state-of-the-art.

Mengze Li, Tianbao Wang, Haoyu Zhang, Shengyu Zhang, Zhou Zhao, Jiaxu Miao, Wenqiao Zhang, Wenming Tan, Jin Wang, Peng Wang, Shiliang Pu, Fei Wu. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Current methods of multi-person pose estimation typically treat the localization and association of body joints separately. In this paper, we propose the first fully end-to-end multi-person Pose Estimation framework with TRansformers, termed PETR. Our method views pose estimation as a hierarchical set prediction problem and effectively removes the need for many hand-crafted modules like RoI cropping, NMS and grouping post-processing. In PETR, multiple pose queries are learned to directly reason a set of full-body poses. Then a joint decoder is utilized to further refine the poses by exploring the kinematic relations between body joints. With the attention mechanism, the proposed method is able to adaptively attend to the features most relevant to target keypoints, which largely overcomes the feature misalignment difficulty in pose estimation and improves the performance considerably. Extensive experiments on the MS COCO and CrowdPose benchmarks show that PETR plays favorably against state-of-the-art approaches in terms of both accuracy and efficiency. The code and models are available at https://github.com/hikvision-research/opera.

Data-free quantization has recently been a promising method to perform quantization without access to the original data. However, the drawback of such approaches is the homogenization of synthetic data due to low efficiency for diverse data generation and the performance collapse of the generator. To alleviate the above issue, we propose a novel Meta-BNS for adversarial data-free quantization scheme which consists of Meta-BNS module and adversarial exploration module. Meta-BNS module automatically learns an enhancement coefficient matrix function for BN loss module to provide a suitable constrain on the generator. Adversarial exploration module leverages minimax game between the generator and quantized model via input gradient to encourage the generator to learn high-dimensional and complex real data distribution. The experimental results show that our method achieves state-of-the-art performance for various settings on data-free quantization.

Feature Pyramid Network (FPN) has been a generic feature extractor in computer vision tasks, which utilizes multi-level features to generate discriminative pyramidal representations. However, the way simply using Sum or Concatenate operation on features to integrate multi-scale information is not sufficient to obtain discriminative semantic representations. In this paper, we propose a dynamic feature pyramid network (DyFPN) to merge multi-scale information in both features and weights. DyFPN uses both high-level context features and low-level spatial structural features to obtain dynamic convolution kernel that contains multi-scale information. In this manner, each resolution in the pyramid performs unique and adaptive convolution directly, meanwhile strengthening the information flow. Specially, DyFPN can be regarded as a complementary enhancement to existing feature pyramid networks. We analyze the effective receptive field and attention map of DyFPN. It proves that our method contains more local information and global information compared with merging multi-scale information only on feature level. Benefit from multi-ways of integrating multi-scale information, our method outperforms other existing feature pyramid methods on COCO detection tasks by a large margin.

Solid results from Transformers have made them prevailing architectures in various natural language and vision tasks. As a default component in Transformers, Layer Normalization (LN) normalizes activations within each token to boost the robustness. However, LN requires on-the-fly statistics calculation in inference as well as division and square root operations, leading to inefficiency on hardware. What is more, replacing LN with other hardware-efficient normalization schemes (e.g., Batch Normalization) results in inferior performance, even collapse in training. We find that this dilemma is caused by abnormal behaviors of activation statistics, including large fluctuations over iterations and extreme outliers across layers. To tackle these issues, we propose Unified Normalization (UN), which can speed up the inference by being fused with other linear operations and achieve comparable performance on par with LN. UN strives to boost performance by calibrating the activation and gradient statistics with a tailored fluctuation smoothing strategy. Meanwhile, an adaptive outlier filtration strategy is applied to avoid collapse in training whose effectiveness is theoretically proved and experimentally verified in this paper. We demonstrate that UN can be an efficient drop-in alternative to LN by conducting extensive experiments on language and vision tasks. Besides, we evaluate the efficiency of our method on GPU. Transformers equipped with UN enjoy about 31% inference speedup and nearly 18% memory reduction. Code will be released at https://github.com/hikvision-research/Unified-Normalization.

Vision Transformers (ViTs) yield impressive performance across various vision tasks. However, heavy computation and memory footprint make them inaccessible for edge devices. Previous works apply importance criteria determined independently by each individual component to prune ViTs. Considering that heterogeneous components in ViTs play distinct roles, these approaches lead to suboptimal performance. In this paper, we introduce joint importance, which integrates essential structural-aware interactions between components for the first time, to perform collaborative pruning. Based on the theoretical analysis, we construct a Taylor-based approximation to evaluate the joint importance. This guides pruning toward a more balanced reduction across all components. To further reduce the algorithm complexity, we incorporate the interactions into the optimization function under some mild assumptions. Moreover, the proposed method can be seamlessly applied to various tasks including object detection. Extensive experiments demonstrate the effectiveness of our method. Notably, the proposed approach outperforms the existing state-of-the-art approaches on ImageNet, increasing accuracy by 0.7% over the DeiT-Base baseline while saving 50% FLOPs. On COCO, we are the first to show that 70% FLOPs of FasterRCNN with ViT backbone can be removed with only 0.3% mAP drop. The code is available at https://github.com/hikvision-research/SAViT.

As it is costly to densely annotate large scale datasets for supervised semantic segmentation, extensive semi-supervised methods have been proposed. However, the accuracy, stability and flexibility of existing methods are still far from satisfactory. In this paper, we propose an effective and flexible framework for semi-supervised semantic segmentation using a small set of fully labeled images and a set of weakly labeled images with bounding box labels. In our framework, position and class priors are designed to guide the annotation network to predict accurate pseudo masks for weakly labeled images, which are used to train the segmentation network. We also propose a mixed-dual-head training method to reduce the interference of label noise while enabling the training process more stable. Experiments on PASCAL VOC 2012 show that our method achieves state-of-the-art performance and can achieve competitive results even with very few fully labeled images. Furthermore, the performance can be further boosted with extra weakly labeled images from COCO dataset.

The accurate segmentation of breast masses in mammography images is a key step in the diagnosis of early breast cancer. To solve the problem of various shapes and sizes of breast masses, this paper proposes a cascaded UNet architecture, which is referred to as CasUNet. CasUNet contains six UNet subnetworks, the network depth increases from 1 to 6, and the output features between adjacent subnetworks are cascaded. Furthermore, we have integrated the channel attention mechanism based on CasUNet, hoping that it can focus on the important feature maps. Aiming at the problem that the edges of irregular breast masses are difficult to segment, a multi-stage cascaded training method is presented, which can gradually expand the context information of breast masses to assist the training of the segmentation model. To alleviate the problem of fewer training samples, a data augmentation method for background migration is proposed. This method transfers the background of the unlabeled samples to the labeled samples through the histogram specification technique, thereby improving the diversity of the training data. The above method has been experimentally verified on two datasets, INbreast and DDSM. Experimental results show that the proposed method can obtain competitive segmentation performance.

Multi-person pose estimation is an attractive and challenging task. Existing methods are mostly based on two-stage frameworks, which include top-down and bottom-up methods. Two-stage methods either suffer from high computational redundancy for additional person detectors or they need to group keypoints heuristically after predicting all the instance-agnostic keypoints. The single-stage paradigm aims to simplify the multi-person pose estimation pipeline and receives a lot of attention. However, recent single-stage methods have the limitation of low performance due to the difficulty of regressing various full-body poses from a single feature vector. Different from previous solutions that involve complex heuristic designs, we present a simple yet effective solution by employing instance-aware dynamic networks. Specifically, we propose an instance-aware module to adaptively adjust (part of) the network parameters for each instance. Our solution can significantly increase the capacity and adaptive-ability of the network for recognizing various poses, while maintaining a compact end-to-end trainable pipeline. Extensive experiments on the MS-COCO dataset demonstrate that our method achieves significant improvement over existing single-stage methods, and makes a better balance of accuracy and efficiency compared to the state-of-the-art two-stage approaches.

Low bit-width model quantization is highly desirable when deploying a deep neural network on mobile and edge devices. Quantization is an effective way to reduce the model size with low bit-width weight representation. However, the unacceptable accuracy drop hinders the development of this approach. One possible reason for this is that the weights in quantization intervals are directly assigned to the center. At the same time, some quantization applications are limited by the various of different network models. Accordingly, in this paper, we propose Multiple Phase Adaptations (MPA), a framework designed to address these two problems. Firstly, weights in the target interval are assigned to center by gradually spreading the quantization range. During the MPA process, the accuracy drop can be compensated for the unquantized parts. Moreover, as MPA does not introduce hyperparameters that depend on different models or bit-width, the framework can be conveniently applied to various models. Extensive experiments demonstrate that MPA achieves higher accuracy than most existing methods on classification tasks for AlexNet, VGG-16 and ResNet.