LRDNet

Please note that at this time (due to a busy schedule and some other reasons), only a brief description is provided. However, one can still access the details and re-generate the results.
This repository contains the official implementation of LRDNet. LRDNet is a lightweight and new method that efficiently detects free road space. The proposed network is lightweight having only 19.5 M parameters (approximately). To date, the LRDNet has the least parameters and the lowest processing time.

Please note that I have cleaned the code, but haven't tested the cleaned code. If there's any problem, please let me know. Moreover, the code is implemented in Keras as we further want it to work on embedded devices. Pytorch despite being faster than Keras on CPU and GPU environments, it seems that Pytorch isn't supported on embedded devices that can achieve up to 300 FPS.

Please refer to the code, the code is self-descriptive. However, here is an overall description of the files:

test.py (Test a trained model. The model names need to be specified in test.py i.e., model_path)
train.py (Code for training. Must specify keywords in the model variable)
trainc.py (Continue train, must specify the weights file that needs to be continued)
AUG  [Folder] (Contains files to generate static augmentations using MATLAB)
ADI  [Folder] (Modified code for ADI) [Current, I'm asking for permission to share the modified code]

Samples results on the KITTI Road Benchmark dataset

FPS and parameter info

Please note that Keras is much much slower than PyTorch. However, as mentioned earlier, we still stick to Keras as we need to port it to embedded devices for future work. The results are presented in the following Table. These results were tested using 2080Ti GPU with 188 GB ram on a 48-core Xeon processor.

Dependencies

Requirements: Tensorflow-gpu==1.14.0, Keras==2.2.4, Tqdm, pillow, NumPy Simply run each code.
For FLOPS:
Download net_flops.py from https://github.com/ckyrkou/Keras and keep it on the root folder (if FLOPS calculation is required
For Backbones
Install Segmentation Models for backbone using instructions from https://github.com/qubvel/segmentation
For Augmentation
Install Albumentations using details from https://github.com/albumentations-team/albumentations or https://albumentations.ai/

How to download dataset(s)?

Please refer to data provider websites
KITTI Road Benchmark
City Scapes
R2D
Place the dataset into data>testing and data>training

For Augmentation data, place the data in data>data_road_aug>train and data>data_road_aug>val

How to covert to BEV (Birds Eye View)?

The resultant results and masks will be stored in seg_results_images folder. As KITTI evaluates the masks in BEV, follow the guideline provided by KITTI road kit.

How to evaluate BEV?

Create an account with your official email and follow the guideline provided by KITTI road kit .

Download pre-trained weights for testing

We provide many pretrained weight files. The models that were used to evaluate on the KITTI evaluation server are represented by their names i.e., LRDNet+, LRDNet(s), and LRDNet(L).

The weight files (HDF5 with learning gradients and abstraction layers preserved, so don’t panic from large file size), the submitted BEV, our modified ADI, and HTML evidence of KITTI submissions can be downloaded from here.

Please cite As:

Currently, the paper is submitted. Further details will be added soon after the response from the journal. No preprint uploaded yet.

abdkhanstd/LRDNet