This example shows an implementation of LaneNet on DepthAI in the Gen2 API system. Model is pretrained on TuSimple dataset. It is taken from PINTO's Model ZOO and converted with model optimizer so that no additional preprocessing is need.
Input video is resized to 512 x 256 (W x H). The inference speed is around 2-4 FPS
Example video is taken from here.
The model produces two outputs - (1) binary segmentation mask where 1 represents line and (2) embeddings of dimension 4 for each pixel. Note that our post-processing differs from the original post-processing in LaneNet, and was made just to showcase the model on the DepthAI.
We currently perform DBSCAN clustering of embeddings, which were sorted using Numpy's lexsort. This works OK and lanes are usually allocated to the same cluster during different frames, as long as no new line is detected on the left side of the leftmost line. For proper post-processing please refer to LaneNet paper.
- Download sample videos.
python3 download.py - Install requirements.
python3 -m pip install -r requirements.txt
python3 main.py [options]
Options:
- -v, --video_path: Path to the video input for inference. Default: vids/vid3.mp4.
- -nn, --nn_model: Select model path for inference. Default: models/lanenet_openvino_2021.4_6shave.blob