Resize the input image

[PeterH0323]

안녕하세요.

Thanks for your great great great job!

I want to resize my image while training , what should I do ?

Please give me some advice or maybe you can add it in the code 😄

[2gunsu]

Hello. Thanks for your interest in my code. Like other papers, this code applies only padding so that the resolution of the image is a multiple of 32, and does not consider resizing at all.

Currently, I am personally implementing the resize transform that can be applied to 3d detection, and I will update the code when the implementation is complete. Thanks again for your interest. 😊

[PeterH0323]

OK, looking forward to it .

감사합니다.

[2gunsu]

Hello. The transform that resizes the image and label has been updated in the code.

The default transforms used in the training are as follows. (Can be found in dataset/monocon_dataset.py)

default_train_transforms = [
    PhotometricDistortion(
        brightness_delta=32,
        contrast_range=(0.5, 1.5),
        saturation_range=(0.5, 1.5),
        hue_delta=18),
    RandomShift(prob=0.5, shift_range=(-32, 32), hide_kpts_in_shift_area=True),
    RandomHorizontalFlip(prob=0.5),
    RandomCrop3D(prob=0.5, crop_size=(320, 960), hide_kpts_in_crop_area=True),
    Normalize(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375]),
    Pad(size_divisor=32),
    ToTensor(),
]

If you want to resize the image to (256, 768), set the first element of the transform list as Resize3D(target_hw=(256, 768)) as shown below. Since the size of the image has changed, when performing a transform that requires a size parameter like RandomCrop3D, you need to set this parameter smaller than the image size.

default_train_transforms = [
    Resize3D(target_hw=(256, 768)),
    PhotometricDistortion(
        brightness_delta=32,
        contrast_range=(0.5, 1.5),
        saturation_range=(0.5, 1.5),
        hue_delta=18),
    RandomShift(prob=0.5, shift_range=(-32, 32), hide_kpts_in_shift_area=True),
    RandomHorizontalFlip(prob=0.5),
    RandomCrop3D(prob=0.5, crop_size=(256, 320), hide_kpts_in_crop_area=True),
    Normalize(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375]),
    Pad(size_divisor=32),
    ToTensor(),
]

Also, if you want to perform resizing during validation or testing, add Resize3D to default_test_transforms as shown below.

default_test_transforms = [
    Resize3D(target_hw=(256, 768)),
    Normalize(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375]),
    Pad(size_divisor=32),
    ToTensor(),
]

Actually, I don't know how the performance will be if the resizing proceeds. 🤔 Have a nice day!

[PeterH0323]

안녕하세요. I will try it. Thx for your great great great job ! 감사합니다 !