add impossible object count by resolution to DG

[Louis-Dupont]

Motivation

Some dataset include huge images with small objects. Resizing to a smaller size may lead to introducing objects that would be too small.

Solution

Plotting a heatmap showing for each image size, how many bbox are smaller than a threshold.

*Feel free to also comment on the plot

[BloodAxe]

Great visualization! I think there is one thing that can make this plot way more informative. Which is some highlight of the image area range in the dataset. Let's say all our samples from data-loader are of 640x640 size. So we should either draw this row with a different color, maybe add annotation or box, etc. The idea is to highlight "where we are now" and how may boxes we are losing currently. If we do not show this information in the chart user would have to either go to the beginning of the report and check image size distribution or look at the code. I think we def. should show some love to users and add this information tip on this plot.

[Louis-Dupont]

Which is some highlight of the image area range in the dataset. Let's say all our samples from data-loader are of 640x640 size. So we should either draw this row with a different color, maybe add annotation or box, etc. The idea is to highlight "where we are now" and how may boxes we are losing currently. If we do not show this information in the chart user would have to either go to the beginning of the report and check image size distribution or look at the code. I think we def. should show some love to users and add this information tip on this plo

I agree on the motivation, but I'm not sure about the solution. This could work when all images are of same size, but if not then it would not be possible. An alternative would be to check the median and show it. But I think that instead of drawing a line, it would be nicer to add an extra row.

[Louis-Dupont]

Update;

What do you think?