Bug during RGB converting in data pipeline

[wonjuleee]

Describe the bug

When creating Image from datumato media, the media has been already converted to RGB as described in both cases

1. Image from path:

   training_extensions/src/otx/api/entities/image.py

   Line 94 in b4fcc6d

   return cv2.cvtColor(cv2.imread(self.__file_path), cv2.COLOR_BGR2RGB)

2. Image from data:

   training_extensions/src/otx/core/data/adapter/base_dataset_adapter.py

   Line 412 in b4fcc6d

   return cv2.cvtColor(data, cv2.COLOR_BGR2RGB)

But this has been again converted to RGB in data pipeline, for instances, in classification tasks:

training_extensions/src/otx/algorithms/classification/configs/base/data/data_pipeline.py

Line 8 in b4fcc6d

__img_norm_cfg = dict(mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)

As a result, there are some performance drop with current data pipelines..

to_rgb	True (as-is)	False (to-be)
CIFAR100 (#2500 training data)	Acc: 0.423 / E2E time: 120.16 sec	Acc: 0.473 / E2E time: 159.17 sec
CUB200 (#1200 training data)	Acc: 0.615 / E2E time: 41.27 sec	Acc: 0.680 / E2E time: 41.92 sec

Steps to Reproduce

1. Turn to_rgb=False
2. Do experiments

Environment:

• OS: Ubuntu 20.04
• Framework version: PyTorch 2.0
• Python version: Python 3.10
• OpenVINO version:
• CUDA/cuDNN version: 11.7
• GPU model and memory: RTX3090

[sovrasov]

If datumaro image is the only source of images in OTX train, we have to make similar changes in all our components. As far as I understand, this change won't affect pytorch models, so we don't have to change OV-related things.

[sungmanc]

Thanks for the reporting, I attached the reason of this issue and made a PR for that.

[goodsong81]

@sungmanc Can we close this issue by #2585 ?

[sungmanc]

@sungmanc Can we close this issue by #2585 ?

Ah, thanks.