visipedia/caltech-fish-counting

The Caltech Fish Counting Dataset

The Caltech Fish Counting Dataset

This repository includes resources for The Caltech Fish Counting Dataset.

• Caltech Fish Counting (CFC)

  • Introduced in The Caltech Fish Counting Dataset: A Benchmark for Multiple-Object Tracking and Counting (ECCV 2022).

• Caltech Fish Counting – Domain Adaptive Object Detection (CFC-DAOD)

  • Introduced in Align and Distill: A Unified Framework for Domain Adaptive Object Detection (2024, Preprint coming soon).

References

@inproceedings{cfc2022eccv,
    author    = {Kay, Justin and Kulits, Peter and Stathatos, Suzanne and Deng, Siqi and Young, Erik and Beery, Sara and Van Horn, Grant and Perona, Pietro},
    title     = {The Caltech Fish Counting Dataset: A Benchmark for Multiple-Object Tracking and Counting},
    booktitle = {European Conference on Computer Vision (ECCV)},
    year      = {2022}
}

@misc{kay2024align,
      title={Align and Distill: Unifying and Improving Domain Adaptive Object Detection}, 
      author={Justin Kay and Timm Haucke and Suzanne Stathatos and Siqi Deng and Erik Young and Pietro Perona and Sara Beery and Grant Van Horn},
      year={2024},
      eprint={2403.12029},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

caltech-fish-counting's Issues

Training and Testing Code

Thank you for sharing the dataset.

Could you please share the model training and testing code? Thanks

Tiny Dataset PAGE NOT FOUND

It seems like the link is broken for the tiny dataset. Could you update it with the correct one?

Mismatched number of frames in annotations vs raw for TinyDataset.

Hi, I'm playing around with the tiny version of this dataset and when I inspected tiny_dataset/annotations-tiny/elwha/Elwha_2018_OM_ARIS_2018_07_09_2018-07-09_190000_490_941/gt_tiny.txt, it seems to contain 47 frames ranging from 208-254. However, when I look at the raw data at tiny_dataset/raw/elwha/Elwha_2018_OM_ARIS_2018_07_09_2018-07-09_190000_490_941, there seems to be 50 frames ranging from 197 to 246. What explains the discrepancy I'm seeing?

mot2yolo.py and coco2yolo.py give differing results (coco2yolo.py CORRECT)

coco2yolo and mot2yolo give differing results.

For example, ...yolo_data/labels/train-v4/2018-05-30-JD150_LeftFar_Stratum1_Set1_LO_2018-05-30_230004_2222_2422_100.txt that was used to train the eccv model gives:

0 0.4097222222222222 0.10032102728731943 0.20138888888888884 0.020866773675762437

After downloading mot annotations from CaltechData, and pulling this latest repo, I run python mot2yolo.py --in_dir ../Data/fishcounting/eccv_2022/annotations_mot/ --out_dir ./labels_from_mot --image_dir ../Data/fishcounting/eccv_2022/frames/raw/

When I look at the same file, i.e. , labels_from_mot/labels/kenai-train/2018-05-30-JD150_LeftFar_Stratum1_Set1_LO_2018-05-30_230004_2222_2422/100.txt I get

0 0.5729166666666666 0.04173354735152488 0.2361111111111111 0.02247191011235955

(I looked at 101.txt to see if it was an off-by-one kind of error - no match)

When I run coco2yolo, python coco2yolo.py --in_dir ../Data/fishcounting/eccv_2022/coco_formatted_annotations --out_dir labels_from_coco
I get

0 0.4097222222222222 0.10032102728731943 0.2013888888888889 0.02086677367576244

coco2yolo agrees with our existing yolo data
mot2yolo does not