search

OpenDriveLab / ViDAR

[CVPR 2024 Highlight] Visual Point Cloud Forecasting
https://arxiv.org/abs/2312.17655
Apache License 2.0
235 stars 15 forks source link

About point cloud forecasting in paper table2 #30

Closed qiuqc1 closed 1 month ago

qiuqc1 commented 1 month ago

ViDAR-RN101-nus-1-8-3future image In the paper, the cd of the input 3s historical frame at CD@1s, CD@2s, and CD@3s does not seem to be consistent with the parameters you gave in the log.

image image It can be found that the table representing ViDAR-RN101-nus-1-8-3future in the logs and the table on github are consistent

tomztyang commented 1 month ago

Use the whole nuScenes dataset to reproduce the reported results. Results on nus-1-8-3future only use 12.5% data in nuScenes.

qiuqc1 commented 1 month ago

image Hello, I got a checkpoint according to ./configs/vidar_pretrain/nusc_fullset/vidar_full_nusc_1future.py, but when I evaluate according to this checkpoint, the final result is only the first frame. What's the situation? image The table corresponding to the 1s time horizons input of the entire data set in the paper is this

tomztyang commented 1 month ago

1 future in 1future_config means we only predict 1 future without future decoder. You can refer to the config and the code for more details.

The 1s history horizon in the paper refers to the model only uses 2 history frames, which refers to this line. 1s history horizon means queue_length=2.

To reproduce Table 2, please modify the 3 future configs according to (./configs/vidar_pretrain/nusc_fullset/vidar_full_nusc_1future.py), using the whole datasets instead.