ALINA: Advanced Line Identification and Notation Algorithm

Paper | Website | CVPR 2024 Presentation

Official implementation of ALINA: Advanced Line Identification and Notation Algorithm [1], accepted to the CVPR 2024 Workshop on Vision Datasets Understanding (VDU).

📦 What's in this Repo

Core pipeline (alina/)

• Interactive ROI selection, drawn once and reused across an entire video's frames
• Perspective warp of the ROI into a bird's-eye view
• HSV color feature normalization + color thresholding
• Histogram-based peak detection to locate candidate line markings
• CIRCLEDAT: the traversal algorithm that isolates line-marking pixels
• Frame unwarping, annotation, and coordinate file output
• Data prep utilities: video → frames, frame rotation, batch resizing

Evaluation (eval/)

• cbem, an interactive tool to hand-create ground truth (context-based edge maps, CBEM)
• evaluate, batch precision / recall / F1 against that ground truth
• superimpose, a visual check overlaying detected coordinates on a Canny edge map

Sample data (data/)

Folder	Contents
Raw_Data/vidd_1, vidd_2, vidd_3	Raw extracted frames from 3 taxiway videos
Labeled_Data/vidd_*/annotations/	Sample ALINA output — labeled frames
Labeled_Data/vidd_*/textfiles/	Sample ALINA output — matching coordinate .txt files
gt_alina_labels/canny_textfiles/	Hand-created CBEM ground-truth coordinates, 3 subsets
gt_alina_labels/canny_images/	CBEM edge-map images (visual reference for the above)
gt_alina_labels/ALINA_textfiles/	ALINA's own output on those same ground-truth frames, for direct comparison

Repository structure

alina/
  config.py       # PipelineConfig / ROIConfig — all tunables in one place
  roi.py          # interactive ROI selection (+ display-scaling for large frames)
  color.py        # HSV color feature normalization
  histogram.py    # vertical projection / peak detection
  traversal.py    # CIRCLEDAT: circular threshold pixel discovery and traversal
  pipeline.py     # process_image() / run_batch() — the main labeling loop
  io_utils.py     # video -> frames, frame rotation, batch resize

eval/
  metrics.py      # precision / recall / F1
  cbem.py         # interactive tool to hand-create ground truth (CBEM)
  evaluate.py     # batch precision/recall/F1 vs. ground truth
  superimpose.py  # visual check: overlay coordinates on a Canny edge map

cli.py            # `alina <subcommand>` entry point
pyproject.toml    # pip install -e .
requirements.txt

data/
  Raw_Data/            # raw extracted video frames (vidd_1, vidd_2, vidd_3)
  Labeled_Data/         # sample ALINA output: annotations/ (images) + textfiles/ (coords)
  gt_alina_labels/      # ground truth used for evaluation
    canny_textfiles/       # CBEM coordinate files (hand-created ground truth)
    canny_images/          # CBEM edge-map images (visual reference)
    ALINA_textfiles/       # ALINA's own output on those same ground-truth frames

Setup

Works with Python 3.9+ (tested with 3.10.12).

git clone <this-repo>
cd ALINA
python3 -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -e .

This installs numpy, opencv-python, matplotlib, and puts an alina command on your PATH (inside the venv). Verify with:

alina --help

Every command below also has an alternative python cli.py ..., run from the repo root.

Running the pipeline

alina label (batch)

Labels every frame in a folder using one ROI, drawn once on the first frame and reused for the rest. Use this for the normal end-to-end workflow: point it at a folder of raw frames from one video, and it labels all of them in one run.

alina label \
  --input-dir data/Raw_Data/vidd_1 \
  --output-images-dir outputs/vidd_1/annotated \
  --output-coords-dir outputs/vidd_1/coords \
  --log-file outputs/vidd_1/timing.log

python cli.py label \
  --input-dir data/Raw_Data/vidd_1 \
  --output-images-dir outputs/vidd_1/annotated \
  --output-coords-dir outputs/vidd_1/coords \
  --log-file outputs/vidd_1/timing.log

Arg	Default	Meaning
--input-dir	–	Folder of .jpg frames to label
--output-images-dir	–	Where annotated frames are written
--output-coords-dir	–	Where per-frame coordinate .txt files are written
--log-file	–	Also write the per-frame log + summary to this file
--peak-threshold	50	Min histogram peak height to attempt line extraction
--circular-threshold	15	CIRCLEDAT neighborhood radius (pixels)
--min-white-pixels	200	Min white pixels in the peak column to trust it
--yellow-lower	0 70 170	Lower HSV bound for the color mask (H S V)
--yellow-upper	255 255 255	Upper HSV bound for the color mask (H S V)
--mask-ignore-left-columns	300	Zero out this many mask columns from the left before traversal

You'll be shown the first frame and asked to click the ROI: Bottom-Left, Top-Left, Top-Right, Bottom-Right, in that order, then press any key. You'll then see a preview of the finalized ROI drawn on the reference frame: press any key to start the batch if it looks right, or Ctrl+C and re-run the command to redo the ROI if not. The selected ROI is applied to every frame in the folder.

Note: Frames larger than 1080px on their longest side are shown at a scaled-down size so the window fits your screen, and clicks are mapped back to full-resolution coordinates automatically.

alina label --image (single frame)

You may use this command on one image from the dataset to tune parameters (--peak-threshold, --circular-threshold, --min-white-pixels, --yellow-lower/--yellow-upper, --mask-ignore-left-columns) before committing to a full batch.

alina label \
  --image data/Raw_Data/vidd_1/00001.jpg \
  --output-images-dir outputs/debug \
  --output-coords-dir outputs/debug \
  --show

python cli.py label \
  --image data/Raw_Data/vidd_1/00001.jpg \
  --output-images-dir outputs/debug \
  --output-coords-dir outputs/debug \
  --show

Arg	Default	Meaning
--image	–	Process a single image instead of a whole --input-dir
--show	off	Display the annotated result window

alina video-to-frames

Extracts every frame of a video file into a folder of numbered .jpg images. Use this first, before labeling, if you're starting from a raw video instead of already-extracted frames.

alina video-to-frames --video raw/taxiway.mp4 --output-dir data/Raw_Data/vidd_4

python cli.py video-to-frames --video raw/taxiway.mp4 --output-dir data/Raw_Data/vidd_4

alina rotate-frames

Rotates every frame in a folder by a fixed angle. Use this if a video was recorded with the camera mounted upside-down or sideways, before running label on it.

alina rotate-frames --input-dir data/Raw_Data/vidd_4 --output-dir outputs/rotated --angle 180

python cli.py rotate-frames --input-dir data/Raw_Data/vidd_4 --output-dir outputs/rotated --angle 180

Arg	Default	Meaning
--angle	180	90, 180, or 270

alina resize-images

Resizes all frames in a folder to a fixed resolution.

alina resize-images --input-dir data/Raw_Data/vidd_4 --output-dir outputs/resized --width 1920 --height 1080

python cli.py resize-images --input-dir data/Raw_Data/vidd_4 --output-dir outputs/resized --width 1920 --height 1080

Arg	Default	Meaning
--width / --height	1920 / 1080	Target resolution

alina cbem (create ground truth)

You can hand-trace a line marking on one frame and run Canny edge detection inside that traced region to produce ground-truth coordinates. Use this to build (or add to) a ground-truth set before running evaluate — it's how the gt_alina_labels/canny_* data in this repo was created.

alina cbem --image data/Raw_Data/vidd_1/00001.jpg --output-dir outputs/cbem

python cli.py cbem --image data/Raw_Data/vidd_1/00001.jpg --output-dir outputs/cbem

Click and drag to trace an outline around a visible line marking, press s to run Canny edge detection inside the traced region and save it, or Esc to close. Multiple regions can be traced and saved from the same frame.

alina evaluate

Computes precision, recall, and F1 by comparing ALINA's coordinate output against ground-truth (CBEM) coordinate files. Use this after you have both a ground-truth set (from cbem) and ALINA's output (from label) for the same frames, to measure how accurate the labeling was.

alina evaluate \
  --canny-dirs data/gt_alina_labels/canny_textfiles/canny_textfiles_1 data/gt_alina_labels/canny_textfiles/canny_textfiles_2 data/gt_alina_labels/canny_textfiles/canny_textfiles_3 \
  --alina-dirs data/gt_alina_labels/ALINA_textfiles/ALINA_textfiles_1 data/gt_alina_labels/ALINA_textfiles/ALINA_textfiles_2 data/gt_alina_labels/ALINA_textfiles/ALINA_textfiles_3

python cli.py evaluate \
  --canny-dirs data/gt_alina_labels/canny_textfiles/canny_textfiles_1 data/gt_alina_labels/canny_textfiles/canny_textfiles_2 data/gt_alina_labels/canny_textfiles/canny_textfiles_3 \
  --alina-dirs data/gt_alina_labels/ALINA_textfiles/ALINA_textfiles_1 data/gt_alina_labels/ALINA_textfiles/ALINA_textfiles_2 data/gt_alina_labels/ALINA_textfiles/ALINA_textfiles_3

Prints average recall, precision, and F1 across all matched frame pairs. --canny-dirs/--alina-dirs are paired by position.

alina superimpose

Draws a given coordinate file's points on top of a Canny edge map of the same frame, so you can visually confirm the detected pixels superimpose the line marking.

alina superimpose \
  --image data/Raw_Data/vidd_1/00001.jpg \
  --coords outputs/vidd_1/coords/00001.txt \
  --output-dir outputs/superimpose_check

python cli.py superimpose \
  --image data/Raw_Data/vidd_1/00001.jpg \
  --coords outputs/vidd_1/coords/00001.txt \
  --output-dir outputs/superimpose_check

--output-dir is optional; omit it to just view the windows without saving.

Benchmark results

ALINA vs. CDLEM [2], evaluated against a 120-frame CBEM ground-truth set:

Algorithm	Detection Rate (%)	Processing Time (ms)
CDLEM	91.14	120.35
ALINA	98.45	50.09

CIRCLEDAT vs. sliding-window search [3], for isolating line-marking pixels:

Algorithm	Time Complexity	Processing Time (ms)
Sliding Window	O(m × n)	10.90
CIRCLEDAT	O(k)	3.33

References

[1] Khan, M. A. H., Ganeriwala, P., Bhattacharyya, S., Neogi, N., & Muthalagu, R. (2024, June). Alina: Advanced line identification and notation algorithm. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) (pp. 7293-7302). IEEE.

[2] Ganeriwala, P., Bhattacharyya, S., Gunther, S., Kish, B., Khan, M. A. H., Dhadoti, A., & Neogi, N. (2023, December). Assisttaxi: A comprehensive dataset for taxiway analysis and autonomous operations. In 2023 International Conference on Machine Learning and Applications (ICMLA) (pp. 1094-1099). IEEE.

[3] Muthalagu, R., Bolimera, A., & Kalaichelvi, V. (2020). Lane detection technique based on perspective transformation and histogram analysis for self-driving cars. Computers & Electrical Engineering, 85, 106653.

@inproceedings{khan2024alina,
  title={Alina: Advanced line identification and notation algorithm},
  author={Khan, Mohammed Abdul Hafeez and Ganeriwala, Parth and Bhattacharyya, Siddhartha and Neogi, Natasha and Muthalagu, Raja},
  booktitle={2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
  pages={7293--7302},
  year={2024},
  organization={IEEE}
}

Acknowledgement

This work is built with OpenCV, NumPy, and Matplotlib, and evaluated on the AssistTaxi dataset [2]. We thank the authors of these projects for making them available to the community.