Skip to content

Releases

PET counter v2: more kinds of crowds

Choose a tag to compare

@khushi-git-7 khushi-git-7 released this 26 Oct 18:17
6acdc13
Merge pull request #39 from khushi-git-7/accuracy-page

Accuracy page: clearer numbers, chart and table fixed

Point model: PET fine-tuned on ShanghaiTech A+B

Choose a tag to compare

@khushi-git-7 khushi-git-7 released this 20 Oct 16:50

PET (Point quEry Transformer, Liu et al., ICCV 2023, MIT licence, academic use) fine-tuned on ShanghaiTech Part A + Part B.

Started from the authors' JHU-Crowd++ model, which never saw ShanghaiTech. Fine-tuned on 90% of the A and B training photos for 9.5 hours on a free Kaggle T4 GPU; the other 10% (70 photos) picked the best epoch (240) and fitted one count correction, true = 1.375 x raw^0.942. The test photos were not used until the final evaluation.

Test results with the correction (accuracy = 1 - total error / total true count; MAE = people per photo):

  • Part A (packed crowds): 86.5% accurate, MAE 58.5 (DM-Count: 83.8%, 70.1)
  • Part B (streets): 92.7% accurate, MAE 9.0 (DM-Count: 86.4%, 16.9)

SHA-256: 5a7c6590e61938e42c6b3a2bbde5fbec34ee7eb8b2c9cae0fdba2a45d8437595

Density model: DM-Count on ShanghaiTech A+B

Choose a tag to compare

@khushi-git-7 khushi-git-7 released this 20 Oct 06:17

DM-Count density model trained on ShanghaiTech Part A + Part B (VGG19 backbone, MIT-licensed code from cvlab-stonybrook/DM-Count).

How it was trained: 90% of the A and B training photos for training, the other 10% held out to pick the best epoch (epoch 130 of 870). The test photos were not used until the final evaluation.

Test results (mean absolute error, people per photo):

  • Part A (packed crowds): 70.1, against 406.9 for the YOLO detector
  • Part B (streets): 16.9, against 73.1 for the YOLO detector

SHA-256: d6f72bffcd1ea986011796b43a63e2738fe8c97d19d230ef99286b603ea207b9