dalzaleq / rgb_lidar_segmentation

Semantic Segmentation for RGB - Lidar data using the following models: Dual SqueezeSeg, Resnet32FCN, EfficientUnet, fcn32, Mobileunet

RGB-Lidar Semantic Segmentation

This code was trained and tested in the following enviorenment

python = 3.6.5 torch = 1.1.0 torchvision = 0.3.0 PIL = 6.1.0 numpy = 1.15.0

Example results

ResFcn_34 Output
Ground truths

Download Data and trained model weights

The data[Lidar + RGB] was converted to PGM representation taking the form of a 64x512 image. There are 8 such channels, XYZ D[Depth] I[Intensity] RGB

This data is stored in npy format, details of the data generations can be found on here

To download data and trained model weights use

./get_data.sh

This will also install the efficientnet_pytorch library with pretrained weights

Remember to change the following in config.py before training:

ROOT_DIR: Data path 
ARGS_ROOT: Path pointing to code directory

Config File

Tune hyper parameters, path files, input channels.

config.py

You only need to change ARGS_ROOT to point to the repo

Model Specific Parameters

ARGS_MODEL_NAME: model name as declared in models/ directory
ARGS_MODEL: directory in which to save best model weights

ARGS_INPUT_TYPE_1: channels that correspond to the input of the first part of the network
ARGS_INPUT_TYPE_2: channels that correspond to the input of the second part of the network

ARGS_TRAIN_BATCH_SIZE: Batch size for training
ARGS_LEARNING_RATE_INIT: Initial Learning rate
ARGS_NUM_EPOCHS: Number of epoch to train the model

Data Stats

To extract stats of data in ROOT_DIR run

python3 calculate_weights.py

after navigating into utils

The following files will be created:

data_stats.ini : which contains the mean and variance of train data for XYZDI
class_weights.ini : which contains the paszke weights for class imbalance

Training models

After making the changes to the config file, run

CUDA_VISIBLE_DEVICES=0 python3 train.py

Inference on Models in the report

Run the following:

python3 inference_table.py

The output should look like this

Inference on trained models

Change the following variables in infer_config.py

ARGS_ROOT: Data path 
ROOT_DIR: Path pointing to code directory

ARGS_MODEL_NAME: model name as declared in models/
ARGS_MODEL: directory in which to save best model weights

ARGS_INPUT_TYPE_1: channels that correspond to the input of the first part of the network
ARGS_INPUT_TYPE_2: channels that correspond to the input of the second part of the network
ARGS_VAL_BATCH_SIZE: Batch size for inference
ARGS_NUM_WORKERS: number of CPU cores to use during evaluation 

To run inference:

python3 infer_single.py