# 2022 / 10 / 18 進度

###### tags: `實驗`
[TOC]

## 瘋狂補實驗
### Classification
1. 判斷是否有需要標註的物件

| 設定 | Best AUC |
| -------- | -------- |
| Unblance data| 0.8204 |
| Random balance data 400x400|0.8607 |
| Whole Img + 8Patch 256x512 | 0.8232 |
| Last Img feature mix 400x400 | 0.8839 |
| Add Last Img FC 400x400 | <font color=red>0.9214</font> |

### Object Detection
1. 測試只用原圖片(1024x1024)3種模型的預測結果

| Model | AP 50 | AP 50-95 | Recall | Time(100epoch) |
| -------- | -------- | -------- | -------- | -------- |
| Faster RCNN |0.2254|0.2750| - |75min|
| FCOS |<font color=red>0.2456</font>|0.2804| - |65min|
| YOLOv7 |0.2173| - |0.3600|100min|

2.Patch-base (要補算Recall)

| Model | AP 50 | AP 50-95 | Recall |
| -------- | -------- | -------- | -------- |
| FCOS |0.2456|0.2804| - |
| Patch-base FCOS |0.5422|0.5982| - |
| Patch-base FCOS RM Sidebox |<font color=red>0.6879</font>|0.6223| - |
| Patch-base FCOS Overlap |0.6304|0.6257| - |

Augmentation 實驗設定有誤，需要重補
| Model | AP 50 | AP 50-95 |
| -------- | -------- | -------- |
| base |0.542|0.6034| - |
| Bright |0.5495|0.5976| - |
| Flip |0.5619|0.6205| - |
| Rotate5 |0.5599|0.6279| - |

### 問題
原本Overlap有兩個問題
1. 當bbox只有切到一點點的時候，根本看不出來也有算框
2. 有跨到下個patch的時候，上界位置

### 雖然排除了以上的問題但overlap的方式還是train的不好mAP都達不到0.6

2022/10/19
使用SGD 1e-3 200epoch 甚至mAP只有0.3