Human Pose estimation-001 model input size mismatch

[Ai-econ]

Model input size mentioned in accuracy-check yaml file was not aligned with model input size mentioned in Readme.
What is actual model input size for the accuracy so that we able to achieve 42% map and is it AP 0.5 or AP 0.5-0.95

Accuracy-check yaml: [https://github.com/openvinotoolkit/open_model_zoo/blob/270dcb5d995b7b5f1db2c08c0db3283a78393e42/models/intel/human-pose-estimation-0001/accuracy-check.yml#L17]

README:[https://github.com/openvinotoolkit/open_model_zoo/blob/270dcb5d995b7b5f1db2c08c0db3283a78393e42/models/intel/human-pose-estimation-0001/README.md?plain=1#L28]

[Wovchena]

The size specified in accuracy-check.yml is actual model input size for the accuracy so that we able to achieve 42% map. README.md states the shape that you get after the model is converted. That's a common "technique" when shape for inference doesn't match the shape for accuracy.

[Ai-econ]

Thanks a lot for quick response

[eaidova]

@Ai-econ for achieving accurate results model required preserving aspect ratio of input image, during accruacy validation model input was reshaped according provided preprocessing rules

[Ai-econ]

@eaidova Where i can find preprocessing rules or guidelines so we can follow the same

If i am not wrong, model was trained with higher resolution which results in good accuracy. However if we deploy the model with lower resolution just by converting model input shape will leads to decrease my accuracy and issue in extracting features right ?

[eaidova]

@Ai-econ they provided in accuracy checker config preprocessing section. Image resized with preserving aspect ratio, where height size will be 368 and width resized according to these rules

open_model_zoo/tools/accuracy_checker/openvino/tools/accuracy_checker/preprocessor/resize.py

Line 33 in 270dcb5

def scale_width(dst_width, dst_height, image_width, image_height,):

and then if image height or width not divisible to 8, it will be additionally padded.

To sum up, resolution used for accuracy evaluation will depends from image size and it is bigger then provided by default (in model, default height - 256, while in accuracy - 368).