annotation data

[habibian]

Hi, could you please share your annotation files? i.e. '/store/datasets/UA-Detrac/COCO-format/test-1-on-30_b.json' or sharing some code elaborating how the annotations are being created? Thanks.

[hu64]

Hi! The script used for my annotations is "object detection/src/tools/convert_csv_to_coco.py". I use csv files with basic format (one object per line, etc.) that themselves are converted from the official XML annotations of UA-DETRAC, which you can find on their website. I can definitely share my files with you, I will put the link here soon.

[hu64]

You can now find them at same location as our weights: https://polymtlca0-my.sharepoint.com/:f:/g/personal/hughes_perreault_polymtl_ca/EhqSkfDIJ-JBh9_YhCrPQrEBocvfP6BIucODKdcNjZzlcA?e=niahaB. Thank you for your interest!

[habibian]

Thanks for your swift response and for sharing the files. The shared folder includes:

• test_b.jason
• train_b.json
• train-1-on10_b.json
• val_b.jason
• val-1-on-10_b.json

I wonder which splits should I use for training and for evaluation to replicate the results reported in your paper?

the default values in the code are (as in UADETRAC1ON10_b): if split == 'test': self.annot_path = '/store/datasets/UA-Detrac/COCO-format/test-1-on-30_b.json' elif split == 'val': self.annot_path = '/store/datasets/UA-Detrac/COCO-format/val_b.json' else: self.annot_path = '/store/datasets/UA-Detrac/COCO-format/train-1-on-10_b.json'

Thank you 👍

[hu64]

Hi, we train on train-1-on10_b.json, monitor on either val_b.json, or val-1-on-10_b.json (faster), and do an evaluation on test-1-on-30_b.json (I will add it). This gives results, but in coco format. To get UA-DETRAC official results, we use the run_on_csv script on the entire test set and do the evaluation on their Matlab benchmark.

[habibian]

Great. Thanks for the clarification :)

[sisrfeng]

Hi, I can not download the files from the links above. Would you mind sharing them in another way or sending them to lwf17@mails.tsinghua.edu.cn or 584400706@qq.com ? Many thanks!

[hu64]

Hi, I just sent the files to both emails using WeTransfer. I hope this works for you! Download link (will expire in one week): https://we.tl/t-ir9etWyY5o

[sisrfeng]

Downloading the full folder from OneDrive or WeTransfer fails. (I am in China) I just need the annotation files so I download them instead of your full folder. Succeed in few minutes. Thank you very much!!

[sisrfeng]

I asked you about ignore_regions some months ago, and I find that handling the ignore_regions improve the results. But I can not find the convert script now. Would you mind sharing your scrip which converts UA-DETRAC to COCO? I want to add the ignore regions to it and share you later.

[hu64]

Hi, my script is in the repo, in src/tools/convert_csv_to_coco.py. The input file is the csv format, one line per detection. the format is: file_name, x0, y0, x1, y1, label.

[sisrfeng]

The official annotations are:

Could you please share your scrpit converting XML files to csv?( I tried this tool but failed: https://github.com/knadh/xmlutils.py)

[hu64]

Yeah sure here it is. You can play around with it, change the validation set and whether the label is the actual label are just 'object' for binary classification. I had not intended to share this so you might have to adapt it to your system. (I joined a .txt file since GitHub won't let me join a Python script, but of course it is a Python script). convert_to_csv.txt

[sisrfeng]

Thank you for your quick reply!