ImageDetRecordIter arbitrary image size

[caiqi]

Hi, it seems that ImageDetRecordIter only supports fix image size now. Is there any way to support arbitrary image size for training SSD with keeping image ratios. Thanks.

[ghost]

@caiqi The default code use a data_shape of 3 x 300 x 300 or 3 x 512 x 512. I've modified it to support arbitrary image size. Here are my modifications:

1. Add multi-scale support for argparse parameters in train.py, evaluate.py, deploy.py and demo.py. Modify the "--data-shape" argument to: parser.add_argument('--data-shape', 'data_shape', type=int, nargs='+', default=[300], help='set image size')

2.Modify the data-shape processing code in train_net.py. In function train_net, line 178: if len(data_shape) ==1: data_shape = (3, data_shape[0], data_shape[0]) elif len(data_shape) == 2: data_shape = (3, data_shape[0], data_shape[1]) else: data_shape = tuple(data_shape)

3. Assuming your input size is 3-channel image WxH, use "--data-shape 3 H W" or "--data-shape H W"; otherwise, if your input size is 1-channel grayscale WxH, use "--data-shape 1 H W".

[caiqi]

@xioryu Thank you very much. I originally thought of getting different size of images in different batches for batch size ==1 like that in Faster R-CNN to resize images with shorter size to 600.