Timezone: »
One-stage detector basically formulates object detection as dense classification and localization (i.e., bounding box regression). The classification is usually optimized by Focal Loss and the box location is commonly learned under Dirac delta distribution. A recent trend for one-stage detectors is to introduce an \emph{individual} prediction branch to estimate the quality of localization, where the predicted quality facilitates the classification to improve detection performance. This paper delves into the \emph{representations} of the above three fundamental elements: quality estimation, classification and localization. Two problems are discovered in existing practices, including (1) the inconsistent usage of the quality estimation and classification between training and inference, and (2) the inflexible Dirac delta distribution for localization. To address the problems, we design new representations for these elements. Specifically, we merge the quality estimation into the class prediction vector to form a joint representation, and use a vector to represent arbitrary distribution of box locations. The improved representations eliminate the inconsistency risk and accurately depict the flexible distribution in real data, but contain \emph{continuous} labels, which is beyond the scope of Focal Loss. We then propose Generalized Focal Loss (GFL) that generalizes Focal Loss from its discrete form to the \emph{continuous} version for successful optimization. On COCO {\tt test-dev}, GFL achieves 45.0\% AP using ResNet-101 backbone, surpassing state-of-the-art SAPD (43.5\%) and ATSS (43.6\%) with higher or comparable inference speed.
Author Information
Xiang Li (NJUST)
Wenhai Wang (Nanjing University)
Lijun Wu (Sun Yat-sen University)
Shuo Chen (Nanjing University of Science and Technology)
Xiaolin Hu (Tsinghua University)
Jun Li (Nanjing University of Science and Technology)
Jinhui Tang (Nanjing University of Science and Technology)
Jian Yang (Nanjing University of Science and Technology)
More from the Same Authors
-
2021 Spotlight: A$^2$-Net: Learning Attribute-Aware Hash Codes for Large-Scale Fine-Grained Image Retrieval »
Xiu-Shen Wei · Yang Shen · Xuhao Sun · Han-Jia Ye · Jian Yang -
2021 Poster: 3D Siamese Voxel-to-BEV Tracker for Sparse Point Clouds »
Le Hui · Lingpeng Wang · Mingmei Cheng · Jin Xie · Jian Yang -
2021 Poster: A$^2$-Net: Learning Attribute-Aware Hash Codes for Large-Scale Fine-Grained Image Retrieval »
Xiu-Shen Wei · Yang Shen · Xuhao Sun · Han-Jia Ye · Jian Yang -
2021 Poster: Learning from Inside: Self-driven Siamese Sampling and Reasoning for Video Question Answering »
Weijiang Yu · Haoteng Zheng · Mengfei Li · Lei Ji · Lijun Wu · Nong Xiao · Nan Duan -
2021 Poster: R-Drop: Regularized Dropout for Neural Networks »
xiaobo liang · Lijun Wu · Juntao Li · Yue Wang · Qi Meng · Tao Qin · Wei Chen · Min Zhang · Tie-Yan Liu -
2021 Poster: Learning to Adapt via Latent Domains for Adaptive Semantic Segmentation »
Yunan Liu · Shanshan Zhang · Yang Li · Jian Yang -
2020 Poster: Generating Adjacency-Constrained Subgoals in Hierarchical Reinforcement Learning »
Tianren Zhang · Shangqi Guo · Tian Tan · Xiaolin Hu · Feng Chen -
2020 Spotlight: Generating Adjacency-Constrained Subgoals in Hierarchical Reinforcement Learning »
Tianren Zhang · Shangqi Guo · Tian Tan · Xiaolin Hu · Feng Chen -
2020 Poster: Causal Intervention for Weakly-Supervised Semantic Segmentation »
Dong Zhang · Hanwang Zhang · Jinhui Tang · Xian-Sheng Hua · Qianru Sun -
2020 Oral: Causal Intervention for Weakly-Supervised Semantic Segmentation »
Dong Zhang · Hanwang Zhang · Jinhui Tang · Xian-Sheng Hua · Qianru Sun -
2019 Poster: Curvilinear Distance Metric Learning »
Shuo Chen · Lei Luo · Jian Yang · Chen Gong · Jun Li · Heng Huang -
2018 Poster: Learning to Teach with Dynamic Loss Functions »
Lijun Wu · Fei Tian · Yingce Xia · Yang Fan · Tao Qin · Lai Jian-Huang · Tie-Yan Liu -
2018 Poster: Designing by Training: Acceleration Neural Network for Fast High-Dimensional Convolution »
Longquan Dai · Liang Tang · Yuan Xie · Jinhui Tang -
2017 Poster: Deliberation Networks: Sequence Generation Beyond One-Pass Decoding »
Yingce Xia · Fei Tian · Lijun Wu · Jianxin Lin · Tao Qin · Nenghai Yu · Tie-Yan Liu