Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adaptive Training Sample Selection

By Shifeng Zhang, Cheng Chi, Yongqiang Yao, Zhen Lei, Stan Z. Li. 2019

2,144 citations recorded by the source corpus.

Object detection has been dominated by anchor-based detectors for several years. Recently, anchor-free detectors have become popular due to the proposal of FPN and Focal Loss. In this paper, we first point out that the essential difference between anchor-based and anchor-free detection is actually how to define positive and negative training samples, which leads to the performance gap between them. If they adopt the same definition of positive and negative samples during training, there is no obvious difference in the final performance, no matter regressing from a box or a point. This shows that how to select positive and negative training samples is important for current object detectors. Then, we propose an Adaptive Training Sample Selection (ATSS) to automatically select positive and negative samples according to statistical characteristics of object. It significantly improves the performance of anchor-based and anchor-free detectors and bridges the gap between them. Finally, we discuss the necessity of tiling multiple anchors per location on the image to detect objects. Extensive experiments conducted on MS COCO support our aforementioned analysis and conclusions. With the newly introduced ATSS, we improve state-of-the-art detectors by a large margin to 50.7% AP without introducing any overhead. The code is available at https://github.com/sfzhang15/ATSS.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

End-to-End Object Detection with Transformers (2020, background): “Their performances are signiﬁcantly inﬂuenced by postprocessing steps to collapse near-duplicate predictions, by the design of the anchor sets and by the heuristics that assign target boxes to anchors [52].”
Swin Transformer: Hierarchical Vision Transformer using Shifted Windows (2021, methodology): “For the ablation study, we consider four typical object detection frameworks: Cascade Mask R-CNN [26, 6], ATSS [71], RepPoints v2 [12], and Sparse RCNN [52] in mmdetection [10].”
Deformable DETR: Deformable Transformers for End-to-End Object Detection (2020, neutral): “es test-time augmentations including horizontal ﬂip and multi-scale testing. Method Backbone TTA AP AP 50 AP 75 AP S AP M AP L FCOS (Tian et al., 2019) ResNeXt-101 44.7 64.1 48.4 27.6 47.5 55.6 ATSS (Zhang et al., 2020) ResNeXt-101 + DCN X 50.7 68.9 56.3 33.2 52.9 62.4 TSD (Song et al., 2020) SENet1”
Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection (2020, methodology): “Specifically, bounding box representation is modeled as a simple Dirac delta distribution [10, 18, 32, 26, 31], which is widely used over past years.”
Dynamic Head: Unifying Object Detection Heads with Attentions (2021, methodology): “We ﬁnd recent studies [12, 35, 28, 31, 33] only focus on solving one of the aforementioned problems in various ways.”
MiniVLM: A Smaller and Faster Vision-Language Model (2020, background): “Region feature is built on top of the object detector, and the detector can be two-stage [38, 8, 24, 55] or one-stage [37, 48, 29, 27, 6, 58, 53, 49].”
YOLOv4: Optimal Speed and Accuracy of Object Detection (2020, methodology): “Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection [94] ATSS ResNet-101 800× 17.”
Focal Self-attention for Local-Global Interactions in Vision Transformers (2021, methodology): “In Table 3 To further verify the effectiveness of our proposed Focal Transformers, we follow [44] to train four different object detectors including Cascade R-CNN [7], ATSS [81], RepPoints [72] and Sparse R-CNN [55].”
PVT v2: Improved baselines with Pyramid Vision Transformer (2021, methodology): “We evaluate Swin Transformer and PVTv2 on four state-of-the-arts detectors, including Cascade R-CNN [1], ATSS [37], GFL [17], and Sparse R-CNN [26].”
Conditional DETR for Fast Training Convergence (2021, methodology): “Typical methods include YOLOv1 [30], CornerNet [21], ExtremeNet [50], CenterNet [49, 6], FCOS [39] and others [23, 28, 52, 19, 51, 22, 15, 46, 47].”
BEVDet: High-performance Multi-camera 3D Object Detection in Bird-Eye-View (2021, methodology): “Two-stage method FasterRCNN [33], one-stage method RetinaNet [20], and their derivatives [10, 2, 4, 39, 49] are the dominant methods in this area [21, 35].”
Generalized Focal Loss V2: Learning Reliable Localization Quality Estimation for Dense Object Detection (2020, methodology): “For many existing approaches [33, 40, 18], they fail to produce the highest LQE scores for the best candidates.”
VarifocalNet: An IoU-aware Dense Object Detector (2020, methodology): “In this paper, we actually study the ATSS version of FCOS (FCOS+ATSS) in which the Adaptive Training Sample Selection (ATSS) mechanism [12] is used to define foreground and background points on the feature pyramids during training.”
Probabilistic two-stage detection (2021, methodology): “While one-stage detectors have achieved competitive performance (Zhang et al., 2020b; Kim & Lee, 2020; Zhang et al., 2019; Li et al., 2020b; Zhu et al., 2020a), they usually rely on heavier separate classiﬁcation and regression branches than two-stage models.”
Sparse R-CNN: End-to-End Object Detection with Learnable Proposals (2020, methodology): ““TTA” indicates test-time augmentations, following the settings in [47].”
YOLOv10: Real-Time End-to-End Object Detection (2024, background): “Over the past years, substantial efforts have been directed towards developing efficient detectors [19, 57, 48, 34, 79, 75, 32, 31, 41].”
YOLOv6: A Single-Stage Object Detection Framework for Industrial Applications (2022, methodology): “Compared with the ATSS, SimOTA can increase AP by 2.0%, and TAL brings 0.5% higher AP than SimOTA.”