Current methods of multi-person pose estimation typically treat the localization and association body joints separately. In this paper, we propose first fully end-to-end Pose Estimation framework with TRansformers, termed PETR. Our method views as a hierarchical set prediction problem effectively removes need for many hand-crafted modules like RoI cropping, NMS grouping post-processing. PETR, multiple queries are learned to directly reason full-body poses. Then joint decoder is utilized...

Fast and precise object detection for hgigh-resolution aerial images has been a challenging task over the years. Due to sharp variations in scale, rotation, aspect ratio, most existing methods are inefficient imprecise. In this paper, we propose different approach polar method. We locate an by centre-point, direct it four angles, measure ratio system. Our coordinate-based method, PolarDet, is faster, simpler, more accurate one-stage detector. Also, our detector introduces sub-pixel centre...

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

Multi-person pose estimation is an attractive and challenging task. Existing methods are mostly based on two-stage frameworks, which include top-down bottom-up methods. Two-stage 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 simplify multi-person pipeline receives a lot of attention. However, recent have limitation low...

Camera and LIDAR are both important sensor modalities for real-world applications, especially autonomous driving. The sensors provide complementary information make fusion possible. However, the progress of early-fusion is very slow due to limitations viewpoint misalignment, feature misalignment data volume alignment, so that its performance also low. In this work, we propose a novel pipeline: an method range image RGB enhance 3D object detection. It takes full advantage LIDAR's view, point...

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 as a direct set prediction problem and effectively removes the need for many hand-crafted components like RoI cropping, one-to-many label assignment, non-maximum suppression (NMS). In multiple queries are learned to directly reason of object embeddings semantic category, bounding-box location, pixel-wise mask in parallel...

3D vehicle detection based on multi-modal fusion is an important task of many applications such as autonomous driving. Although significant progress has been made, we still observe two aspects that calls for further improvement: First, what extra information can be obtained from the images to complement point clouds in tasks seldom explored by previous works. Second, most modules only used their designed network, lacking universality. In this work, propose PointAttentionFusion and...

It is still challenging to detect and locate anomalies by models trained only with normal samples. Methods using image reconstruction as a pretext task can provide precise localization but suffer from harnessing the capability on unseen anomalies. This paper proposes new framework of Multi-Task Hard example Mining (MTHM) for anomaly detection localization. The self-supervised multi-task setting creatively takes advantage competition among different tasks learn more compact efficient...

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 of arbitrary unseen without any re-training. Specifically, the proposed method draws idea from point cloud registration and resorts object-agnostic features establish 3D-3D correspondences between object-scene object-model cloud. Then parameters solved by...

Most existing knowledge distillation methods for semantic segmentation focus on extracting various sophisticated from raw features. However, such is usually manually designed and relies prior as in traditional feature engineering. In this paper, we aim to propose a simple effective method using To end, revisit the pioneering work distillation, FitNets, which simply minimizes mean squared error (MSE) loss between teacher student Our experiments show that naive yields good results, even...

Performance of current point cloud-based outdoor 3D object detection relies heavily on large-scale high-quality annotations. However, such annotations are usually expensive to collect and scenes easily accumulate massive unlabeled data containing rich scenes. Semi-supervised learning is a effective alternative utilize both labeled data, but remains unexplored in detection. Inspired by indoor semi-supervised methods, SESS 3DIoUMatch, we propose ATF-3D, framework for Specifically, design...

The benign and malignant (BM) classification of breast masses based on mammography is a key step in the diagnosis early cancer an effective way to improve survival rate patients. Nevertheless, due differences size, shape texture visual similarity between same category, it difficult obtain robust model using conventional deep learning methods. To address this problem, we proposed Multi-Tasking U-shaped Network (MT-UNet), which contains three ideas: 1) architecture constructed can well adapt...

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 present study, i.e. clean and data decoupling, classification localization task decoupling. In first two-level thresholds used to categorize pseudo into three groups, backgrounds, foregrounds foregrounds. With specially designed noise-bypass head focusing on data, backbone networks can extract coarse but...

Post-training quantization (PTQ) is an effective compression method to reduce the model size and computational cost. However, quantizing a into low-bit one, e.g., lower than 4, difficult often results in non-negligible performance degradation. To address this, we investigate loss landscapes of quantized networks with various bit-widths. We show that network more ragged surface, easily trapped bad local minima, which mostly appears quantization. A deeper analysis indicates, surface caused by...

Phosphodiesterase 5 (PDE5) is a cGMP-specific hydrolytic enzyme and widely distributed in versatile tissues. PDE5 has been identified as valid therapeutic target for treating erectile dysfunction pulmonary arterial hypertension (PAH). Herein, hit-to-lead structural optimizations were performed on the PDE1 inhibitor

In recent years, significant progress has been made on the research of crowd counting. However, as challenging scale variations and complex scenes existed in crowds, neither traditional convolution networks nor Transformer architectures with fixed-size attention could handle task well. To address this problem, paper proposes a scene-adaptive network, termed SAANet. First all, we design deformable in-built backbone, which learns adaptive feature representations sampling locations dynamic...

3D vehicle detection based on multi-modal fusion is an important task of many applications such as autonomous driving. Although significant progress has been made, we still observe two aspects that need to be further improvement: First, the specific gain camera images can bring seldom explored by previous works. Second, algorithms run slowly, which essential for with high real-time requirements(autonomous driving). To this end, propose end-to-end trainable single-stage feature adaptive...

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 simple yet effective framework Distill Knowledge from VLM DETR-like detector, termed DK-DETR. Specifically, present two ingenious distillation schemes named semantic knowledge (SKD) and relational (RKD). To utilize rich systematically, SKD transfers explicitly, while RKD exploits implicit relationship information between...

The accurate segmentation of breast masses in mammography images is a key step the diagnosis early cancer. To solve problem various shapes and sizes masses, this paper proposes cascaded UNet architecture, which referred to as CasUNet. CasUNet contains six subnetworks, network depth increases from 1 6, output features between adjacent subnetworks are cascaded. Furthermore, we have integrated channel attention mechanism based on CasUNet, hoping that it can focus important feature maps. Aiming...