Decreased accuracy when training with non-square inputs

[ghost]

Hello, thank you for the great work.

I have conducted training on both rectangular and square size inputs on my custom dataset. However, the average accuracy of the rectangular trained model is significantly lower. Below is the configuration that was used for training & evaluation:

Square TrainReader:

batch_transforms:
    - BatchRandomResize: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}

Rect TrainReader:

batch_transforms:
    - BatchRandomResize: {target_size: [352, 608], random_size: False, random_interp: True, keep_ratio: False}

Square EvalReader:

EvalReader:
  sample_transforms:
    - Decode: {}
    - Resize: {target_size: [608, 608], keep_ratio: False, interp: 2} # target_size: (h, w)

Rect EvalReader:

EvalReader:
  sample_transforms:
    - Decode: {}
    - Resize: {target_size: [352, 608], keep_ratio: False, interp: 2} # target_size: (h, w)

I've read through #13, but it doesn't seem to be related to training with rectangular inputs. When 'random_size: True,' one of the numbers in the target_size is used for square resizing.
I also tried switching w and h as suggested, but the results were similar.
Could I get any help? I would like to know if there's something I'm doing wrong or if a similar issue has been encountered before.

[lyuwenyu]

1. I think your case is same as 关于数据增强与长方形输入 #13. The target_size: [[800, 960]] is a list including one element [800, 960]. It always random choise same rect size.

  batch_transforms:
    - BatchRandomResize: {target_size: [[800, 960]], random_size: True, random_interp: True, keep_ratio: False}

2. Perhaps your can try the BatchRandomResize style as 关于数据增强与长方形输入 #13 using latest code, although I think the logic is same.

# reset eval_size
eval_size: [352, 608]

  batch_transforms:
    - BatchRandomResize: {target_size: [[352, 608]], random_size: True, random_interp: True, keep_ratio: False}
    - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
    - NormalizeBox: {}
    - BboxXYXY2XYWH: {}
    - Permute: {}

3. Make sure self.inputs['im_shape'] == self.inputs['image'])[2:] in here