From a5816de889c84d4f7fe75c08321a33d1d9ec7ae1 Mon Sep 17 00:00:00 2001
From: pk_hk <82303451+pkhk-1@users.noreply.github.com>
Date: Fri, 11 Nov 2022 15:08:16 +0800
Subject: [PATCH] [ppyoloe-sod] add im_size to box_distribution (#7308)

---
 configs/smalldet/README.md | 12 +++++++++---
 tools/box_distribution.py  | 12 ++++++++++++
 2 files changed, 21 insertions(+), 3 deletions(-)

diff --git a/configs/smalldet/README.md b/configs/smalldet/README.md
index 79a92385e..a5babf569 100644
--- a/configs/smalldet/README.md
+++ b/configs/smalldet/README.md
@@ -136,7 +136,8 @@ Tower(59)
 `
 
 </details>
-，原始数据集[下载链接](https://challenge.xviewdataset.org/download-links)。
+
+，原始数据集[下载链接](https://challenge.xviewdataset.org/)。
 
 
 ### 用户自定义数据集准备
@@ -289,10 +290,15 @@ CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inferenc
 python tools/box_distribution.py --json_path ../../dataset/DOTA/annotations/train.json --out_img box_distribution.jpg
 ```
 - `--json_path` ：待统计数据集COCO 格式 annotation 的json文件路径
+- `--eval_size` ：推理尺度（默认640）
+- `--small_sride` ：模型最小步长（默认8）
 - `--out_img` ：输出的统计分布图路径
 
 以DOTA数据集的train数据集为例，统计结果打印如下：
 ```bash
+Suggested reg_range[1] is 13
+Mean of all img_w is 2304.3981547196595
+Mean of all img_h is 2180.9354151880766
 Median of ratio_w is 0.03799439775910364
 Median of ratio_h is 0.04074914637387802
 all_img with box:  1409
@@ -301,8 +307,8 @@ Distribution saved as box_distribution.jpg
 ```
 
 **注意:**
-- 当原始数据集全部有标注框的图片中，**有1/2以上的图片标注框的平均宽高与原图宽高比例小于0.04时**，建议进行切图训练。
-
+- `Suggested reg_range[1]` 为数据集在优化后DFL算法中推荐的`reg_range`上限，即` reg_max+1`。
+- 当原始数据集全部有标注框的图片中，**原图宽高均值大于1500且有1/2以上的图片标注框的平均宽高与原图宽高比例小于0.04时**，建议进行切图训练。
 
 ### SAHI切图
 
diff --git a/tools/box_distribution.py b/tools/box_distribution.py
index 974936b67..f7979ecc1 100644
--- a/tools/box_distribution.py
+++ b/tools/box_distribution.py
@@ -53,6 +53,7 @@ def get_ratio_infos(jsonfile, out_img, eval_size, small_stride):
     be_im_h = []
     ratio_w = []
     ratio_h = []
+    im_wid,im_hei=[],[]
     for ann in tqdm(allannjson['annotations']):
         if ann['iscrowd']:
             continue
@@ -63,6 +64,8 @@ def get_ratio_infos(jsonfile, out_img, eval_size, small_stride):
         else:
             im_w = coco.imgs[be_im_id]['width']
             im_h = coco.imgs[be_im_id]['height']
+            im_wid.append(im_w)
+            im_hei.append(im_h)
             im_m_w = np.mean(be_im_w)
             im_m_h = np.mean(be_im_h)
             dis_w = im_m_w / im_w
@@ -72,9 +75,16 @@ def get_ratio_infos(jsonfile, out_img, eval_size, small_stride):
             be_im_id = ann['image_id']
             be_im_w = [w]
             be_im_h = [h]
+        
 
     im_w = coco.imgs[be_im_id]['width']
     im_h = coco.imgs[be_im_id]['height']
+    im_wid.append(im_w)
+    im_hei.append(im_h)
+    all_im_m_w = np.mean(im_wid)
+    all_im_m_h = np.mean(im_hei)
+
+
     im_m_w = np.mean(be_im_w)
     im_m_h = np.mean(be_im_h)
     dis_w = im_m_w / im_w
@@ -100,6 +110,8 @@ def get_ratio_infos(jsonfile, out_img, eval_size, small_stride):
     ratio_w = [i * 1000 for i in ratio_w]
     ratio_h = [i * 1000 for i in ratio_h]
     print(f'Suggested reg_range[1] is {reg_max+1}' )
+    print(f'Mean of all img_w is {all_im_m_w}')
+    print(f'Mean of all img_h is {all_im_m_h}') 
     print(f'Median of ratio_w is {mid_w}')
     print(f'Median of ratio_h is {mid_h}')
     print('all_img with box: ', len(ratio_h))
-- 
GitLab