support ICDAR detect

Author: ming71

.gitignore

@@ -4,14 +4,10 @@
 __pycache__/
 *.py[cod]
 
-/dataset/
+
 /runs/
-/data/samples/
-/data/train/
-/data/test/
-/data/all/
-/data/extra/
-/backup/
+/data/HRSC+/
+/data/ICDAR/test/
 /output/
 /weights/
 /utils/kmeans/

LR.png

@@ -606,7 +606,7 @@ activation=linear
 # 角度定义为x+ 0; 顺时针; 正负0.5pi
 [yolo]
 mask = 144-215
-anchors = utils/kmeans/anchor-cluster.txt
+anchors = utils/kmeans/hrsc_512.txt
 classes=1
 num=9
 jitter=.3
@@ -692,7 +692,7 @@ activation=linear
 
 [yolo]
 mask = 72-143
-anchors = utils/kmeans/anchor-cluster.txt
+anchors = utils/kmeans/hrsc_512.txt
 classes=1
 num=9
 jitter=.3
@@ -779,7 +779,7 @@ activation=linear
 
 [yolo]
 mask = 0-71
-anchors = utils/kmeans/anchor-cluster.txt
+anchors = utils/kmeans/hrsc_512.txt
 classes=1
 num=9
 jitter=.3

cfg/HRSC+/yolov3_512.cfg

@@ -662,7 +662,7 @@ activation=linear
 # 角度定义为x+ 0; 顺时针; 正负0.5pi
 [yolo]
 mask = 144-215
-anchors = utils/kmeans/anchor-cluster.txt
+anchors = utils/kmeans/hrsc_512.txt
 classes=1
 num=9
 jitter=.3
@@ -748,7 +748,7 @@ activation=linear
 
 [yolo]
 mask = 72-143
-anchors = utils/kmeans/anchor-cluster.txt
+anchors = utils/kmeans/hrsc_512.txt
 classes=1
 num=9
 jitter=.3
@@ -835,7 +835,7 @@ activation=linear
 
 [yolo]
 mask = 0-71
-anchors = utils/kmeans/anchor-cluster.txt
+anchors = utils/kmeans/hrsc_512.txt
 classes=1
 num=9
 jitter=.3

cfg/HRSC+/yolov3_512_se.cfg

@@ -0,0 +1,42 @@
+giou: 0.1       # giou loss gain 1.582
+cls: 27.76      # cls loss gain  (CE=~1.0, uCE=~20)
+cls_pw: 1.446   # cls BCELoss positive_weight
+obj: 20.35      # obj loss gain (*=80 for uBCE with 80 classes)
+obj_pw: 3.941   # obj BCELoss positive_weight
+iou_t: 0.3      # iou training threshold
+ang_t: 3.1415926/12
+reg: 1.0
+fl_gamma: 0.5   # focal loss gamma
+context_factor: 1.0 # 按照短边h来设置的,wh的增幅相同； 调试时设为倒数直接检测
+
+
+# lr
+lr0: 0.00008
+multiplier:10
+warm_epoch:5
+lrf: -4.        # final LambdaLR learning rate = lr0 * (10 ** lrf)
+momentum: 0.97  # SGD momentum
+weight_decay: 0.0004569  # optimizer weight decay
+
+
+# aug
+hsv_s: 0.5      # image HSV-Saturation augmentation (fraction)
+hsv_v: 0.3      # image HSV-Value augmentation (fraction)
+degrees: 50.0    # image rotation (+/- deg)
+translate: 0.2  # image translation (+/- fraction)
+scale: 0.2      # image scale (+/- gain)
+shear: 0.0
+gamma: 0.2
+blur:  1.2
+noise: 0.005
+contrast: 0.0
+sharpen: 0.0
+copypaste: 0.0  
+grayscale: 0.05 
+
+
+# training
+epochs: 1500
+batch_size: 4
+save_interval: 50
+test_interval: 1