Merge pull request #2 from nuguziii/develop

Develop

Author: nuguziii

.gitignore

@@ -39,6 +39,7 @@ eggs/
 parts/
 sdist/
 var/
+experiments/
 *.egg-info/
 .installed.cfg
 *.egg

experiments/pano/pano_hrnet48.yaml

@@ -0,0 +1,127 @@
+AUTO_RESUME: false
+CUDNN:
+  BENCHMARK: true
+  DETERMINISTIC: false
+  ENABLED: true
+DATA_DIR: ''
+GPUS: (0,)
+OUTPUT_DIR: 'output'
+LOG_DIR: 'log'
+WORKERS: 4
+PRINT_FREQ: 10
+
+DATASET:
+  COLOR_RGB: true
+  DATASET: 'pano'
+  DATA_FORMAT: jpg
+  FLIP: false
+  NUM_JOINTS_HALF_BODY: 0
+  PROB_HALF_BODY: 0.0
+  ROOT: 'data/pano/'
+  ROT_FACTOR: 0
+  SCALE_FACTOR: 0.0
+  TEST_SET: 'test'
+  TRAIN_SET: 'train'
+MODEL:
+  INIT_WEIGHTS: true
+  NAME: pose_hrnet
+  NUM_JOINTS: 32
+  PRETRAINED: ''
+  TARGET_TYPE: gaussian
+  IMAGE_SIZE:
+  - 512
+  - 256
+  HEATMAP_SIZE:
+  - 128
+  - 64
+  SIGMA: 1
+  EXTRA:
+    PRETRAINED_LAYERS:
+      - 'conv1'
+      - 'bn1'
+      - 'conv2'
+      - 'bn2'
+      - 'layer1'
+      - 'transition1'
+      - 'stage2'
+      - 'transition2'
+      - 'stage3'
+      - 'transition3'
+      - 'stage4'
+    FINAL_CONV_KERNEL: 1
+    STAGE2:
+      NUM_MODULES: 1
+      NUM_BRANCHES: 2
+      BLOCK: BASIC
+      NUM_BLOCKS:
+        - 4
+        - 4
+      NUM_CHANNELS:
+        - 48
+        - 96
+      FUSE_METHOD: SUM
+    STAGE3:
+      NUM_MODULES: 4
+      NUM_BRANCHES: 3
+      BLOCK: BASIC
+      NUM_BLOCKS:
+        - 4
+        - 4
+        - 4
+      NUM_CHANNELS:
+        - 48
+        - 96
+        - 192
+      FUSE_METHOD: SUM
+    STAGE4:
+      NUM_MODULES: 3
+      NUM_BRANCHES: 4
+      BLOCK: BASIC
+      NUM_BLOCKS:
+        - 4
+        - 4
+        - 4
+        - 4
+      NUM_CHANNELS:
+        - 48
+        - 96
+        - 192
+        - 384
+      FUSE_METHOD: SUM
+LOSS:
+  USE_TARGET_WEIGHT: true
+TRAIN:
+  BATCH_SIZE_PER_GPU: 16
+  SHUFFLE: true
+  BEGIN_EPOCH: 0
+  END_EPOCH: 200
+  OPTIMIZER: adam
+  LR: 0.001
+  LR_FACTOR: 0.1
+  LR_STEP:
+  - 150
+  - 180
+  WD: 0.0001
+  GAMMA1: 0.99
+  GAMMA2: 0.0
+  MOMENTUM: 0.9
+  NESTEROV: false
+TEST:
+  BATCH_SIZE_PER_GPU: 1
+  COCO_BBOX_FILE: 'data/coco/person_detection_results/COCO_val2017_detections_AP_H_56_person.json'
+  BBOX_THRE: 1.0
+  IMAGE_THRE: 0.0
+  IN_VIS_THRE: 0.2
+  MODEL_FILE: ''
+  NMS_THRE: 1.0
+  OKS_THRE: 0.9
+  USE_GT_BBOX: false
+  FLIP_TEST: false
+  POST_PROCESS: true
+  SHIFT_HEATMAP: false
+DEBUG:
+  DEBUG: true
+  SAVE_BATCH_IMAGES_GT: true
+  SAVE_BATCH_IMAGES_PRED: true
+  SAVE_HEATMAPS_GT: true
+  SAVE_HEATMAPS_PRED: true

lib/config/default.py

@@ -37,7 +37,7 @@
 _C.MODEL.NAME = 'pose_hrnet'
 _C.MODEL.INIT_WEIGHTS = True
 _C.MODEL.PRETRAINED = ''
-_C.MODEL.NUM_JOINTS = 17
+_C.MODEL.NUM_JOINTS = 32
 _C.MODEL.TAG_PER_JOINT = True
 _C.MODEL.TARGET_TYPE = 'gaussian'
 _C.MODEL.IMAGE_SIZE = [256, 256]  # width * height, ex: 192 * 256
@@ -82,6 +82,7 @@
 _C.TRAIN.NESTEROV = False
 _C.TRAIN.GAMMA1 = 0.99
 _C.TRAIN.GAMMA2 = 0.0
+_C.TRAIN.LOSS_WEIGHT = [1e1,1e1,1e1]
 
 _C.TRAIN.BEGIN_EPOCH = 0
 _C.TRAIN.END_EPOCH = 140

lib/core/evaluate.py

@@ -28,7 +28,7 @@ def calc_dists(preds, target, normalize):
     return dists
 
 
-def dist_acc(dists, thr=0.5):
+def dist_acc(dists, thr=0.2):
     ''' Return percentage below threshold while ignoring values with a -1 '''
     dist_cal = np.not_equal(dists, -1)
     num_dist_cal = dist_cal.sum()
@@ -70,4 +70,49 @@ def accuracy(output, target, hm_type='gaussian', thr=0.5):
         acc[0] = avg_acc
     return acc, avg_acc, cnt, pred
 
+def accuracy_classification(output, target, thres=0.0):
+    '''
+    Calculate accuracy according to PCK,
+    but uses ground truth heatmap rather than x,y locations
+    First value to be returned is average accuracy across 'idxs',
+    followed by individual accuracies
+    '''
+
+    output[output<thres] = 0
+    output[output>=thres] = 1
+
+    cnt = output.shape[0]
+    acc = np.zeros(cnt)
+
+    target = target.reshape(cnt, -1)
+
+    for i in range(cnt):
+        acc[i] = sum(output[i]==target[i])/32
+
+    avg_acc = np.mean(acc)
+    return avg_acc, cnt
+
+def accuracy_landmark(output, target, thres=6.0):
+    '''
+    Calculate accuracy according to PCK,
+    but uses ground truth heatmap rather than x,y locations
+    First value to be returned is average accuracy across 'idxs',
+    followed by individual accuracies
+    '''
+    batch = output.shape[0]
+    acc = np.zeros(batch)
+
+    output = output.reshape(batch, 32, 2)
+    target = target.reshape(batch, 32, 2)
+
+    diff = np.sqrt(np.square(output[:,:,0] - target[:,:,0]) + np.square(output[:,:,1] - target[:,:,1]))
+
+    for i in range(batch):
+        cur = diff[i]
+        cur[cur < thres] = 0
+        acc[i] = (32-np.count_nonzero(cur))/32
+
+    avg_acc = np.mean(acc)
+    return avg_acc, batch
+