support train motchallenge and crowdhuman

configs/detection/fcos/fcos_r50_torch_1x_mot20_crowdhuman.py

@@ -0,0 +1,42 @@
+_base_ = './fcos_r50_torch_1x_coco.py'
+
+data_root0 = 'data/tracking/crowdhuman/'
+data_root1 = 'data/tracking/MOT20/'
+CLASSES = ('pedestrian', )
+train_dataset = dict(
+    data_source=dict(
+        ann_file=[
+            data_root1 + 'annotations/train_cocoformat.json', data_root0 +
+            '/annotations/crowdhuman_train.json', data_root0 +
+            '/annotations/crowdhuman_val.json'
+        ],
+        img_prefix=[
+            data_root1 + 'train', data_root0 + 'train', data_root0 + 'val'
+        ],
+        classes=CLASSES))
+
+val_dataset = dict(
+    data_source=dict(
+        ann_file=data_root0 + '/annotations/crowdhuman_val.json',
+        img_prefix=data_root0 + 'val',
+        classes=CLASSES))
+
+data = dict(
+    imgs_per_gpu=2, workers_per_gpu=2, train=train_dataset, val=val_dataset)
+
+model = dict(head=dict(num_classes=1))
+
+optimizer = dict(lr=0.001)
+
+eval_pipelines = [
+    dict(
+        mode='test',
+        evaluators=[
+            dict(type='CocoDetectionEvaluator', classes=CLASSES),
+        ],
+    )
+]
+
+checkpoint_config = dict(interval=1)
+
+load_from = 'https://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/EasyCV/modelzoo/detection/fcos/fcos_epoch_12.pth'

easycv/datasets/builder.py

@@ -8,22 +8,22 @@
 
 
 def _concat_dataset(cfg, default_args=None):
-    ann_files = cfg['ann_file']
-    img_prefixes = cfg.get('img_prefix', None)
-    seg_prefixes = cfg.get('seg_prefix', None)
-    proposal_files = cfg.get('proposal_file', None)
+    ann_files = cfg['data_source']['ann_file']
+    img_prefixes = cfg['data_source'].get('img_prefix', None)
+    seg_prefixes = cfg['data_source'].get('seg_prefix', None)
+    proposal_files = cfg['data_source'].get('proposal_file', None)
 
     datasets = []
     num_dset = len(ann_files)
     for i in range(num_dset):
         data_cfg = copy.deepcopy(cfg)
-        data_cfg['ann_file'] = ann_files[i]
+        data_cfg['data_source']['ann_file'] = ann_files[i]
         if isinstance(img_prefixes, (list, tuple)):
-            data_cfg['img_prefix'] = img_prefixes[i]
+            data_cfg['data_source']['img_prefix'] = img_prefixes[i]
         if isinstance(seg_prefixes, (list, tuple)):
-            data_cfg['seg_prefix'] = seg_prefixes[i]
+            data_cfg['data_source']['seg_prefix'] = seg_prefixes[i]
         if isinstance(proposal_files, (list, tuple)):
-            data_cfg['proposal_file'] = proposal_files[i]
+            data_cfg['data_source']['proposal_file'] = proposal_files[i]
         datasets.append(build_dataset(data_cfg, default_args))
 
     return ConcatDataset(datasets)
@@ -35,6 +35,8 @@ def build_dataset(cfg, default_args=None):
     elif cfg['type'] == 'RepeatDataset':
         dataset = RepeatDataset(
             build_dataset(cfg['dataset'], default_args), cfg['times'])
+    elif isinstance(cfg['data_source'].get('ann_file'), (list, tuple)):
+        dataset = _concat_dataset(cfg, default_args)
     else:
         dataset = build_from_cfg(cfg, DATASETS, default_args)
 

tools/prepare_data/crowdhuman2coco.py

@@ -0,0 +1,102 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import argparse
+import json
+import os
+import os.path as osp
+from collections import defaultdict
+
+import mmcv
+from PIL import Image
+from tqdm import tqdm
+
+# python tools/convert_datasets/crowdhuman2coco.py -i /apsarapangu/disk4/easycv_nfs/data/tracking/crowdhuman -o /apsarapangu/disk4/easycv_nfs/data/tracking/crowdhuman/annotations
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(
+        description='CrowdHuman to COCO Video format')
+    parser.add_argument(
+        '-i',
+        '--input',
+        help='root directory of CrowdHuman annotations',
+    )
+    parser.add_argument(
+        '-o',
+        '--output',
+        help='directory to save coco formatted label file',
+    )
+    return parser.parse_args()
+
+
+def load_odgt(filename):
+    with open(filename, 'r') as f:
+        lines = f.readlines()
+    data_infos = [json.loads(line.strip('\n')) for line in lines]
+    return data_infos
+
+
+def convert_crowdhuman(ann_dir, save_dir, mode='train'):
+    """Convert CrowdHuman dataset in COCO style.
+
+    Args:
+        ann_dir (str): The path of CrowdHuman dataset.
+        save_dir (str): The path to save annotation files.
+        mode (str): Convert train dataset or validation dataset. Options are
+            'train', 'val'. Default: 'train'.
+    """
+    assert mode in ['train', 'val']
+
+    records = dict(img_id=1, ann_id=1)
+    outputs = defaultdict(list)
+    outputs['categories'] = [dict(id=1, name='pedestrian')]
+
+    data_infos = load_odgt(osp.join(ann_dir, f'annotation_{mode}.odgt'))
+    for data_info in tqdm(data_infos):
+        img_name = osp.join('Images', f"{data_info['ID']}.jpg")
+        img = Image.open(osp.join(ann_dir, mode, img_name))
+        width, height = img.size[:2]
+        image = dict(
+            file_name=img_name,
+            height=height,
+            width=width,
+            id=records['img_id'])
+        outputs['images'].append(image)
+
+        if mode != 'test':
+            for ann_info in data_info['gtboxes']:
+                bbox = ann_info['fbox']
+                if 'extra' in ann_info and 'ignore' in ann_info[
+                        'extra'] and ann_info['extra']['ignore'] == 1:
+                    iscrowd = True
+                else:
+                    iscrowd = False
+                ann = dict(
+                    id=records['ann_id'],
+                    image_id=records['img_id'],
+                    category_id=outputs['categories'][0]['id'],
+                    vis_bbox=ann_info['vbox'],
+                    bbox=bbox,
+                    area=bbox[2] * bbox[3],
+                    iscrowd=iscrowd)
+                outputs['annotations'].append(ann)
+                records['ann_id'] += 1
+        records['img_id'] += 1
+
+    if not osp.isdir(save_dir):
+        os.makedirs(save_dir)
+    mmcv.dump(outputs, osp.join(save_dir, f'crowdhuman_{mode}.json'))
+    print(f'-----CrowdHuman {mode} set------')
+    print(f'total {records["img_id"] - 1} images')
+    if mode != 'test':
+        print(f'{records["ann_id"] - 1} pedestrians are annotated.')
+    print('-----------------------')
+
+
+def main():
+    args = parse_args()
+    convert_crowdhuman(args.input, args.output, mode='train')
+    convert_crowdhuman(args.input, args.output, mode='val')
+
+
+if __name__ == '__main__':
+    main()