HotSight / PseudoLabeling /GENERATED_LABELS.md
hotsight's picture
Fix relative paths after the PseudoLabeling/ move and stale split-file names in docs
490858f verified
|
Raw History Blame Contribute Delete
3.7 kB

Automatically generated labels

This PseudoLabeling/ folder holds everything related to automatically generated labels. The hand-labeled dataset it refers to is ../Data/, one level up, and the dataset scripts are in ../scripts/.

Two folders of per-pixel labels produced by the annotation pipeline in annotation_tool/: a trained YOLO detector and IR hotspot thresholding propose candidate regions, and SAM turns those proposals into per-pixel masks, in the same format as the dataset's own masks. The pipeline can be run on other registered RGB+IR footage as well; see its own README. The labels are generated rather than hand-annotated, so spot-check mask_overlay/ before relying on them.

Folder What it contains Use it for Frames
PseudoDataset10s/ (~29 GB) New frames that are not in ../Data/, extracted from the original videos, plus their labels Training augmentation: blend into training data 2,214 frames, 33 clips
GeneratedAnnotations10s/ (~10 GB) Labels for the existing ../Data/ frames (train + test) Reference and comparison only: human labels for these exact frames already exist, so training on them is not recommended 775 frames, 31 clips

Which frames the pseudo dataset adds

../Data/ samples one frame every 300 video frames (10 s at 29.97 fps). PseudoDataset10s fills in between those samples, at offsets +50, +150, +250 of each gap, roughly 3x the original frame count on the same clips. Sampling used ../scripts/sample_dataset.py, so the RGB/IR pairing rule matches the original dataset's.

Each pseudo frame's IR-to-RGB homography is a distance-weighted average of the homographies of the two neighbouring ../Data/ frames. Frames were excluded where those two disagree by more than 20 px, since the drone was moving there and the averaged mapping would be unreliable. Only clips with hand-labeled frames in ../Data/ are covered.

PseudoDataset10s/pseudo_samples.csv has one row per new frame, with the clip, the frame number, and the two frames it was interpolated between. Its split column carries the clip's train/val/test assignment, so filter on it before training.

Layout

Both folders follow the same <Unit>/<Clip>/ tree as ../Data/, with the same fNNNNN frame naming: frames from ../Data/ are multiples of 300, pseudo frames end in 50, 150 or 250.

  • mask/: the label mask, one ASCII digit per pixel, values 1-5, row-major, 2688 x 1512, byte-identical in format to the masks in ../Data/.

  • mask_overlay/, mask_color/: the labels drawn over the RGB frame, and as a flat color image.

  • rgb/, ir/, ir_gray/: the frame pair, with the de-colorized IR following the same convention as ../Data/.

  • homography/: the IR-to-RGB registration for the frame.

Classes are 1 = red, 2 = orange, 3 = yellow, 4 = green, 5 = black. Every pixel gets one of these five; there is no "unlabeled" value.

Contents of this folder

  • PseudoDataset10s/: automatically generated labels for new frames sampled between the 10-second samples.

  • GeneratedAnnotations10s/: automatically generated labels for the frames already in ../Data/, for reference rather than training.

  • annotation_tool/: the labeling pipeline that produced both; see annotation_tool/README.md.

  • generate_pseudo_dataset.py: the script that produced PseudoDataset10s. It picks new frame numbers between the existing ones, extracts each RGB/IR pair, averages the registration from the neighbouring frames, and runs the pipeline on the result. Running it needs the source videos and the annotators' samples_splits.csv, neither of which is part of this release.