Personally I'm a big fan of DotDotGoose [0]. Used by lots of people all over. I discovered this when I had to test accuracy in a pose estimation model for a co

Personally I'm a big fan of DotDotGoose [0]. Used by lots of people all over. I discovered this when I had to test accuracy in a pose estimation model for a computer vision project involving crowds being counted by CCTV. Turns out in low-res imaging (even the best CCTV is a bit rubbish at long range due to wide angle lenses and so on), pose estimation models beat almost everything else for counting. Except... they're still not accurate. I would spend hours going through hundreds of frames, counting people by hand, and then I'd compare with different pose models, and found they were always out, but they were out by the same amount. Some models got confused by reflective surfaces, so would double count. Others would be out by ~20% every time for various environmental factors. The good thing about this is you could then easily show calibration. "OK, in that space, whatever the model says, half it, that's a better count", or "OK, in this image from this camera whatever the number the model says, increase/decrease it by 20%, and that'll be more accurate". We ended up through human calibration being able to provide much more accurate "models" per camera than the models themselves. It did mean counting hundreds of people per frame for hundreds of frames for dozens of cameras though... I had to abandon the trackpad for a mouse and just got settled in for a few days with DotDotGoose. Colleagues were surprised I had the patience for it. :-) https://biodiversityinformatics.amnh.org/open_source/dotdotgoose/ https://biodiversityinformatics.amnh.org/open_source/dotdotg...