Questions in converting cityscapes dataset

[jfzhuang]

Thanks for your excellent work and kindly code releasing. I have a problem when reproducing the cityscapes experiment.

Following your tutorial, I have reproduced the experimental results, 44.37 vs. 51.73 with 6.79 improvement under the cutmix setting. When delving deep into the codes, I notice that you implement a downsample_label_img function for downsampling ground truth in the convert_cityscapes.py. However, it is more common to use nearest downsampling method in cv2 or PIL. Then, I replace the downsample_label_img function with the following:
y_img = cv2.resize(y_img, (1024, 512), interpolation=cv2.INTER_NEAREST)
and re-conduct the cityscapes experiment. The results degrades greatly, 43.79 vs. 47.02 with only 3.23 improvement.

I wonder the reason why a different downsample method affects the result greatly and the motivation to reimplement a downsample method rather than using a common one.

[Britefury]

Hi, thank you for exploring this. I have to admit I didn't give it much thought. When downsampling an RGB image you tend to get better results - at least visually - by using a local mean approach, so I tried to replicate this for labels as best I could. I didn't consider alternatives.

Out of curiosity, do your results arise from a single run or from the average of multiple runs?

[jfzhuang]

Thanks for your quick reply and resolve my concern. The results arise from a single run.

[Britefury]

Sure thing.

You may find that taking the average of a few runs will close the gap between the results from various downsampling methods.