search

hasanirtiza / Pedestron

[Pedestron] Generalizable Pedestrian Detection: The Elephant In The Room. @ CVPR2021
https://openaccess.thecvf.com/content/CVPR2021/papers/Hasan_Generalizable_Pedestrian_Detection_The_Elephant_in_the_Room_CVPR_2021_paper.pdf
Apache License 2.0
689 stars 158 forks source link

Can you provide the code to convert the original annotations of the ECP dataset into COCO format (including vis_ratio)? #172

Closed eorroot closed 4 months ago

eorroot commented 4 months ago

Thank you very much for the json file that converts the annotations of the ECP dataset into COCO format. I am a bit confused that the annotations for the ECP dataset do not provide exact information on vis_ratio and occ, only approximate occ information,, but I see that the Pedestron/datasets/EruoCity/day_val_visT.json file you provided contains explicit vis_ratio information.

hasanirtiza commented 4 months ago

This is the script we used to convert the annotations, (tools/convert_datasets/eurocity/convert_eurocity_to_coco.py). On quick look, doesn't appear to me that we extracted exact vis ratio.

eorroot commented 4 months ago

This is the script we used to convert the annotations, (tools/convert_datasets/eurocity/convert_eurocity_to_coco.py). On quick look, doesn't appear to me that we extracted exact vis ratio. 1720656965588 image

Thank you very much for your reply.

Maybe you misunderstood what I meant. What I mean is that I found two kinds of comments (day_val.json, day_val_visT.json) in (Pedestron/datasets/EuroCity/) about ECP validation sets (I provided the images for your convenience).

I took a quick look at (tools/convert_datasets/eurocity/convert_eurocity_to_coco.py) and it seems that the conversion is to the day_val.json version. And the conversion to day_val_visT.json (which contains the vis_ratio) has no corresponding script . Maybe, you forgot to update it.

hasanirtiza commented 4 months ago

Now it is clear, these file were added by another researcher @AbdulHannanKhan to the repo, please contact him on how he computed this vis ratio (potentially could be through segmentation, or may be he contacted the benchmark publishers).