new post

_posts/journal/2024-01-16-day24.md

@@ -0,0 +1,27 @@
+---
+layout: post
+title: Day24
+date: 2024-01-16 20:13 +0800
+category: [读研日记]
+tags: []
+img_path: "/assets/img/posts/"
+---
+
+花了一周把yolov8的所有scale以及相应的cbam+p234的改进版都跑了一遍
+
+FLIR, 200 epochs
+
+|Model            |Parameters(M)|mAP@50:95(%)|mAP50(%) |
+|:---:            |:---:        |:---:       |:---:    |
+|yolov8n          |3.157        |29.14       |48.07    |
+|yolov8n-cbam-p234|2.334        |31.98       |52.14    |
+|yolov8s          |11.167       |34.71       |55.21    |
+|yolov8s-cbam-p234|7.785        |35.75       |57.25    |
+|yolov8m          |25.903       |37.51       |59.44    |
+|yolov8m-cbam-p234|19.112       |38.40       |59.11    |
+|yolov8l          |43.692       |39.11       |60.38    |
+|yolov8l-cbam-p234|35.671       |39.48       |60.64    |
+|yolov8x          |68.230       |39.04       |60.31    |
+|yolov8x-cbam-p234|55.482       |39.36       |60.10    |
+
+可以看到m, l, x的改进对精度影响极其小，甚至出现了负增长，后续尽量使用s进行实验