- Lane Matching presentation: https://docs.google.com/presentation/d/1_aKsyaCRaU2LQY1fK3FoJUxReD7GSdPmB6vxaPzUNW0/edit?usp=sharing
- Lane Matching ipynb notebook: https://github.com/KomajzCz/FEL-matching-notebook
- main: branch has an old code. It should be similiar with heuristic but README for docker is in common/pipeline. Main purpose is overview of tasks given by Patrik.
- heuristic: actual branch with old segmentation, but it is in final state and can be packaged into Docker
- xyzi_nn_segmentation : actual branch with current version of cylydric segmentation that can be packaged into Docker
- cylinder_segmentation: old branch Ondra than switched to xyzi_nn_segmentation
- dev: very old branch nothing is there
- General: Potential datasets for using existing annotations, images to LiDAR projection annotation, HD maps with lanes and LiDAR
- Person: Ondra
- Tasks:
- List all potentially useful datasets with instance-level annotations
- Make a material (table, half page of text) we can decide on and send to Valeo
- References:
- OpenDriveLab/OpenLane-V2: [NeurIPS 2023 Track Datasets] - No LiDAR
- Argoverse 2 HD maps
- K-LANE
- Input: Search on internet and try to map existings datasets to our needs
- Output:
- List of datasets with parameters in text or table (has id annotations, scene diversity, LiDAR sensor, ...)
- Conclussion on what datasets we can use for learning the model to detect lanes id/polylines
- General: Prepare annotation tool and understand the metrics Lane extraction scenario (K-Lane should use the most common evaluation protocols)
- Person: Honza + Martin
- Tasks:
- Download K-Lane devkit and run annotation GUI
- Annotate one point cloud from Valeo dataset to learn how to use it, annotate full line, not segments (dashed)
- Think about transfer of K-Lane annotations into Valeo polylines format
- Construct meaningful metrics which we can use for evaluation given what we have in Valeo and what we can annotate
- Reference:
- Input: Codebase, existing data from Valeo
- Output: System for annotating data, Exact and systematic protokol, how we should evaluate our results
- General: Download and learn how to use the data and easily transfer with Valeo format
- Person: Yana
- Tasks:
- Download K-Lane dataset
- Learn how to use the dataset
- Convert to Valeo format
- References:
- Input: K-Lane Dataset, Valeo Dataset
- Output: Functions to allow for training on both datasets, merging the formats/annotations
- General: To get baseline for lane detection
- Person: -
- Tasks:
- Learn how to infer the K-Lane model used in Paper
- Run The model on Valeo Dataset
- Calculate metrics on Valeo Dataset
- Retrain models on Valeo Dataset and show metrics
- References:
- Input: Existing codebase, data samples
- Output: Importable models, calculated metrics on both datasets using training on K-Lane and both.
- Camera propagated to LiDAR - To demonstrate additional data gathering
- Argoverse Dataset HD maps - to get additional data
- Pseudo-labelling - to easily boost performance
- Test-Time Augmentation - to easily boost performance
- Active learning framework - to save annotations