Yzichen
Why are there a large number of blank areas in the visualization results？
The configuration file I used is flashocc-r50.py, the weight file is flashocc-r50-256x704.pth, and the dataset is the nuscenes-mini dataset. Run command:
bash tools/dist_test.sh projects/configs/flashocc/flashocc-r50.py ckpts/flashocc-r50-256x704.pth 4 --eval map
python tools/analysis_tools/vis_occ.py --root_path /data/nuscenes/v1.0-mini --save_path ./vis ./results
The visualization results obtained are shown in the following figure.
Are these large blank areas in the visualization results due to missed detections, or is there a problem with my dataset or the visualization? Because when I commented out the display of detection results in the visualization code vis_occ.py and only displayed the ground truth (gt), I found that the gt and the predicted results were almost exactly the same, both showing large blank areas.Is there a feasible approach to address this issue?
Also, why are only the scene-0103 and scene-0916 scenes selected each time? Is there any way to switch to different scenes? @Yzichen Thank you very much for your reply.
You are correct in observing that there are a lot of blank areas in the gt data that should require human correction. Therefore, the model sometimes learns such erroneous results. This may need to be addressed by starting with cleaning or optimising the dataset.
As for the 2nd question, you could try increasing vis-frames?
