BEVNet

Source code for our work "Semantic Terrain Classification for Off-Road Autonomous Driving"

TODOs

Datasets

Datasets should be put inside data/. For example, data/semantic_kitti_4class_100x100.

Download links

SemanticKITTI: Google Drive
RELLIS: Google Drive

Running the pretrained models

Model weights

SemanticKITTI

Single-frame: Google Drive
Recurrent: Google Drive

RELLIS

Single-frame: Google Drive
Recurrent: Google Drive

Run the models

First extract the model weights

cd /path/to/bevnet/experiments
unzip /path/to/zip/file

To run the models on the validation set, cd to bevnet/bevnet, then run

# Single-frame model
python test_single.py --model_file ../experiments/kitti4_100/single/include_unknown/default-logs/model.pth.4 --test_env kitti4

# Recurrent model
python test_recurrent.py --model_file ../experiments/kitti4_100/recurrent/include_unknown/default-logs/model.pth.2 --test_env kitti4

Training

BEVNet-S

Example:

cd experiments
bash train_kitti4-unknown_single.sh kitti4_100/single/include_unknown/default.yaml <tag> arg1 arg2 ...

Logs and model weights will be stored in a subdirectory of the config file like this: experiments/kitti4_100/single/include_unknown/default-<tag>-logs/

<tag> is useful when you want to use the same config file but different hyperparameters. For example, if you want to do some debugging you can use set <tag> to debug.
arg1 arg2 ... are command line arguments supported by train_single.py. For example, you can pass --batch_size=4 --log_interval=100, etc.

BEVNet-R

The command line formats are the same as BEVNet-S Example:

cd experiments
bash train_kitti4-unknown_recurrent.sh kitti4_100/recurrent/include_unknown/default.yaml <tag> \
--n_frame=6 --seq_len=20 --frame_strides 1 10 20 \
--resume kitti4_100/single/include_unknown/default-logs/model.pth.4 \
--resume_epoch 0

Logs and model weights will be stored in a subdirectory of the config file experiments/kitti4_100/recurrent/include_unknown/default-<tag>-logs/.

djramakrishna/semantic_bevnet