Skip to content

Latest commit

 

History

History

chain_interp_flow_preprocess

README.md

Pre-processing for Chained Flow Grouping Experiments

This folder contains files for pre-processing data for chained-flow grouping experiments. Files include scripts for off-line estimation of 3D scene flow, and scripts for interpolation and chaining of 3D scene flow for further frames.

Download Point Cloud Data

Synthia

For Synthia experiments, process the raw Synthia dataset or download the processed point clouds following instructions in here (total size ~5.1GB). Then extract the files and place it like

../semantic_seg_synthia/processed_pc/
    SYNTHIA-SEQS-01-DAWN-000000.npz
    SYNTHIA-SEQS-01-DAWN-000001.npz
    ...
    SYNTHIA-SEQS-06-SUNSET-000841.npz

Off-line Estimation of 3D Scene Flow for Consecutive Frames

In order to off-line estimate 3D scene flow for consecutive frames, download the pre-trained FlowNet3D model from this repo, i.e. download model file from here and checkpoints from here. Make sure log_train/ and model_concat_upsa.py are in this directory.

Synthia

For Synthia experiments, run command_pred_pairwise_flow_synthia.sh by

sh command_pred_pairwise_flow_synthia.sh

The output scene flow will be stored in ../semantic_seg_synthia/init_flow.

Interpolate and Chain Scene Flow for Further Frames

The default number of nearest neighbors is set to be 2. One may change it to try with different settings.

Synthia

For Synthia experiments, run command_process_flow_to_chain_synthia.sh by

sh command_process_flow_to_chain_synthia.sh

The output scene flow will be stored in ../semantic_seg_synthia/chained_flow.