search

mileyan / pseudo_lidar

(CVPR 2019) Pseudo-LiDAR from Visual Depth Estimation: Bridging the Gap in 3D Object Detection for Autonomous Driving
https://mileyan.github.io/pseudo_lidar/
MIT License
976 stars 216 forks source link

Evaluation with pre-trained frustum pointnet #41

Open Anirudh-Swaminathan opened 4 years ago

Anirudh-Swaminathan commented 4 years ago

I have a huge memory error when evaluating with pre-trained frustum pointnet. I downloaded the PseudoLiDAR pointclouds from this repo, as well as the pre-trained FrustumPointNet model. I used a conda environment to have Python 2.7 to ensure working compatibility with the Frustum PointNet repository. I'm using an Ubuntu 18.04 server with 1 GPU.

For the evaluation of Frustum-PointNet, they make it necessary to generate a pickle file of the data. When running their script for generating that, my program gets killed for OOM. On further inspection, I find that while the pickle file for the original LiDAR data is 996MB, for just the first 1000 samples, my frustum_carpedcyc_train.pickle file produces 4.8GB.

The input_list in case of the original data has a shape (1526, 4) for the first data sample, while the input_list in case of PseudoLiDAR data has a shape of (18604, 4). This means that the veoldyne pointclouds are huge in PseudoLiDAR in my understanding.

It would be great if the authors could provide details on how they exactly evaluated on the PseudoLiDAR dataset, or provide the pickle files for the same so that we need not generate those huge files.

s56207824inc commented 3 years ago

@Anirudh-Swaminathan same problem, did you sovle it?