This repository contains the sources of ASSN, which improved the small object dectection performance. This includes the experimental results and the trained models on YOLOv7 and YOLOv8 based on ssFPN, a previous study conducted on YOLOv4 and YOLOR.
This repository was created based on the official repository source of YOLOv7 and YOLOv8.
| Model | Test Size | AP | AP50 | AP75 | APS | APM | APL |
|---|---|---|---|---|---|---|---|
| YOLOv7 | 640 | 51.2% | 69.7% | 55.5% | 35.2% | 56.0% | 66.7% |
| YOLOv7-X | 640 | 52.9% | 71.1% | 57.5% | 36.9% | 57.7% | 68.6% |
| YOLOv7-W6 | 1280 | 54.6% | 72.3% | 59.5% | 40.1% | 59.0% | 68.6% |
| YOLOv8n | 640 | 37.4% | 52.9% | 40.3% | 18.6% | 41.0% | 53.5% |
| YOLOv8s | 640 | 44.9% | 62.1% | 48.3% | 25.9% | 49.9% | 61.0% |
| ASSN + YOLOv7 | 640 | 51.5% | 70.0% | 56.1% | 35.7% | 56.3% | 65.8% |
| ASSN + YOLOv7-X | 640 | 53.5% | 71.6% | 58.2% | 36.7% | 58.3% | 69.7% |
| ASSN + YOLOv7-W6 | 1280 | 55.0% | 72.7% | 60.1% | 40.0% | 59.5% | 68.2% |
| ASSN + YOLOv8n | 640 | 37.4% | 53.2% | 40.5% | 20.5% | 41.6% | 51.6% |
| ASSN + YOLOv8s | 640 | 45.0% | 62.3% | 49.1% | 27.2% | 50.3% | 59.7% |
| Model | Test Size | AP | AP50 | AP75 | APS | APM | APL |
|---|---|---|---|---|---|---|---|
| YOLOv7 | 640 | 51.4% | 69.7% | 55.9% | 31.8% | 55.5% | 65.0% |
| YOLOv7-X | 640 | 53.1% | 71.2% | 57.8% | 33.8% | 57.1% | 67.4% |
| YOLOv7-W6 | 1280 | 54.9% | 72.6% | 60.1% | 37.3% | 58.7% | 67.1% |
| ASSN + YOLOv7 | 640 | 51.9% | 70.3% | 56.5% | 32.6% | 55.7% | 65.4% |
| ASSN + YOLOv7-X | 640 | 53.5% | 71.5% | 58.2% | 34.3% | 57.5% | 67.4% |
| ASSN + YOLOv7-W6 | 1280 | 55.2% | 72.9% | 60.5% | 37.9% | 58.8% | 67.5% |
- Test Set (test-dev2017, 20k images) have an issue on experimental server (https://codalab.lisn.upsaclay.fr/competitions/7384) with YOLOv8, so we cannot get any experimental results with this set for YOLOv8 currently.
| CPU | Intel® Core™ i9-9900K CPU @ 3.60GHz × 16 |
| GPU | NVIDIA GeForce GTX 1080 Ti (11GB) |
| OS | Ubuntu 22.04.3 LTS |
