SelectiveFocus

This is the repo of "Selective Focus: Investigating Semantics Sensitivity in Post-training Quantization for Lane Detection", which is accepted by AAAI 2024. We would publish the pre-print version (containing the supplementary, intending for a better understanding) and the source code.

Source Code

We built this system on other excellent projects:

• mmdetection: https://github.com/open-mmlab/mmdetection
• condlane: https://github.com/aliyun/conditional-lane-detection
• ganet: https://github.com/Wolfwjs/GANet
• pytorch-auto-drive: https://github.com/voldemortX/pytorch-auto-drive
• qdrop: https://github.com/wimh966/QDrop

1. install mmdet

cd mmdet ; python setup.py develop

2. prepare the TuSimple and CurveLane datasets according to PAD. It is worthy note that the BezierLaneNet needs extra data-preprocessing as it adopted a curve-based modeling of lanes. The model checkpoints could be find at PAD, CondlaneNet, GANet.

Also remember to compile the unified evaluate tool on evaluate/culane/ by

make

3. run

we provide scripts to run different quantization settings in the dir scripts and the network and quantization configs are located at dir config.

To run PTQ on PAD[https://github.com/voldemortX/pytorch-auto-drive] models:

python tools/pad/main_lanedet_quant.py \
    --config $network_cfg \
    --quant_config $quant_cfg \
    --test \
    --output_prefix $output_dir \
    2>&1 | tee -a $log_file_name

To run PTQ on CondlaneNet and GANet:

python tools/mmdet/culane_ganet_quant.py \
    $network_cfg \
    $quant_cfg \
    $ckpt_pth_file \
    --result_dst $output_dir \
    --launcher none 2>&1 | tee -a $log_file_name

NOTE: every pair of (dataset, network) gets a single scripts in the tools/.

Supplementary

An arxiv version is on going...