Knowledge Vault 5 /56 - CVPR 2020
Bridging the Gap Between Anchor-based and Anchor-free Detection via Adaptive Training Sample Selection
Shifeng Zhang; Cheng Chi; Yongqiang Yao; Zhen Lei; Stan Li
< Resume Image >

Concept Graph & Resume using Claude 3 Opus | Chat GPT4o | Llama 3:

graph LR classDef atss fill:#f9d4d4, font-weight:bold, font-size:14px classDef detection fill:#d4f9d4, font-weight:bold, font-size:14px classDef experiments fill:#d4d4f9, font-weight:bold, font-size:14px classDef released fill:#f9f9d4, font-weight:bold, font-size:14px A[Bridging the Gap
Between Anchor-based and
Anchor-free Detection via
Adaptive Training Sample
Selection] --> B[ATSS bridges anchor-based and
anchor-free detection. 1] A --> C[CNN detectors: anchor-based
and anchor-free. 2] C --> D[RetinaNet, FCOS differ in
samples, regression, anchors. 3] A --> E[Positive/negative definition essential,
regression starting unimportant. 4] A --> F[ATSS adaptively selects samples,
improves performance. 5] F --> G[ATSS: robust hyperparameter,
insensitive to anchors. 6] F --> H[Multiple anchors may be
unnecessary with ATSS. 7] H --> I[Role of multiple anchors
needs further study. 8] A --> J[Released codes on
authors homepage. 9] A --> K[Explore differences: anchor-based
and anchor-free detection. 10] class A,B,K atss class C,D detection class E,F,G,H,I experiments class J released


1.- ATSS (Adaptive Training Sample Selection) bridges the gap between anchor-based and anchor-free object detection methods.

2.- Current CNN-based object detectors include anchor-based (one-stage and two-stage) and anchor-free (center-based and keypoint-based) methods.

3.- RetinaNet (anchor-based) and FCOS (anchor-free) have similar frameworks but differ in training sample definition, regression starting status, and anchor tiling.

4.- Experiments show that the definition of positive and negative samples is an essential difference, while regression starting status is not.

5.- ATSS adaptively selects positive and negative samples based on statistical characteristics of objects, improving performance without overhead.

6.- ATSS has one robust hyperparameter and is insensitive to anchor settings.

7.- Tiling multiple anchors per location may not be useful after applying an appropriate sample selection strategy like ATSS.

8.- The right role of tiling multiple anchors per location needs further study.

9.- The authors provide released codes on the first author's homepage for others to use their methods.

10.- The work aims to explore the essential differences between anchor-based and anchor-free object detection methods.

Knowledge Vault built byDavid Vivancos 2024