ENPM673 Spring 2022: Perception for Autonomous Robots
Project 4
Instructions -
Semantic segmentation using RGB, LiDAR and Optical flow sensor data Trained on vkitti dataset.
Download vkitti_1.3.1_depthgt.tar, vkitti_1.3.1_flowgt.tar, vkitti_1.3.1_rgb.tar, vkitti_1.3.1_scenegt.tar
from this link. Extract
the tarballs into one folder. Use the path of this folder for <path_to_dataset>
argument throughout the following
instruction.
To train, run the following commands:
python main.py --dataset_path <path_to_dataset> [--lidar] [--optical_flow]
The --lidar
and --optical_flow
flags are optional. Specifying the flags will train the network with corresponding
LiDAR and oflow data. Without the flags, the network will be trained only on RGB data.
To test, run the following commands:
python recons.py --dataset_path <path_to_dataset> --checkpoint_path <path_checkpoint_file> [--lidar] [--optical_flow]
Specify one or both of --lidar
and --optical_flow
flags if you had specified it during training. A pretrained checkpoint file
can be downloaded from here. Note that if you're using
our checkpoint file, do not specify the --lidar
and --optical_flow
flags.
The below image shows the results of FusionNet. The pictures on the left column are the actual segmentation ground truth and the pictures on the right column are the segmentation masks obtained from our net.
For training and testing RGB and Optical flow data follow the instrcutions on the ReadMe of BiSeNet Download initial training weights from *here
To fuse segmentation results from optical flow or rgb with lidar points, modify the input data here
Convert lidar.bin
files to .pcd
using the function
Run the following command after setting all the data
python3 lidar_fusion.py
The output after execution will look something like this
- Raft: optical flow: learnopencv
- Raft: optical flow: paperswithcode
- Flownet 2.0: optical flow
- ppr: SegFlow: Segmentation + optical flow
- git: SegFlow: Segmentation + optical flow
- KITTI Scene Flow dataset
Authors -
Naitri Rajyaguru (117361919)
Mayank Joshi (117555264)
Nitesh Jha (117525366)
Aneesh Dandime (118418506)
Tanuj Thakkar (117817539)