search

bradyz / cross_view_transformers

Cross-view Transformers for real-time Map-view Semantic Segmentation (CVPR 2022 Oral)
MIT License
525 stars 80 forks source link

Argoverse2 Dataset Implementation #46

Open aapatni opened 1 year ago

bradyz commented 1 year ago

hey @aapatni this looks awesome! thanks for working on this. have you tried training yet? We had tried argoverse quite a while ago as well. IIRC the diversity and size of dataset was much smaller than nuScenes, so we just settled on that

aapatni commented 1 year ago

Hey @bradyz , thanks!

Just got finished on a smaller scale training (150 scenes, 5 epochs) and here are the results. Working on using a larger subset of the scene data + more epochs currently.

pred_sampled

Screenshot 2023-04-04 at 5 28 43 PM

One difference between nuscenes and argoverse2 datasets is that argoverse2 doesn't have the concept of 'visibility' which I believe helps the model understand what objects in the scene are visible to a percentage of the cameras. Currently, we just set the visibility to 100% for all objects in the scene.

aoyanl commented 3 months ago

Hey @bradyz , thanks!

Just got finished on a smaller scale training (150 scenes, 5 epochs) and here are the results. Working on using a larger subset of the scene data + more epochs currently.

pred_sampled pred_sampled

Screenshot 2023-04-04 at 5 28 43 PM

One difference between nuscenes and argoverse2 datasets is that argoverse2 doesn't have the concept of 'visibility' which I believe helps the model understand what objects in the scene are visible to a percentage of the cameras. Currently, we just set the visibility to 100% for all objects in the scene.

Wonderful job! I'm trying the argoverse2 recently. What is the iou of vehicle you get? Looking for your reply!