hotsight commited on
Commit
490858f
·
verified ·
1 Parent(s): 4464c5a

Fix relative paths after the PseudoLabeling/ move and stale split-file names in docs

Browse files

Docs and scripts inside PseudoLabeling/ now refer to ../Data/, ../scripts/, ../markdown/ and ../../html/. GENERATED_LABELS.md restored to its full version with these fixes. generate_pseudo_dataset.py defaults to ../Data. Data/README.md and scripts/README.md name the released splits_graminoid.csv / splits_canopy.csv and annotation_label_by_split_*.csv files. samples_splits.csv is documented as not part of the release.

Data/README.md CHANGED
@@ -12,7 +12,7 @@
12
  └── homography/{homography.json,points.json}
13
  ```
14
 
15
- `fXXXXX` is the original zero-based RGB frame index. RGB images are 2688×1512 and IR images are 640×360. Each mask is a flattened 2688×1512 row-major array of class labels `1`–`5`. `samples.csv` and `splits.csv` reports which samples are sampled and belong to which split.
16
 
17
  ## Known gaps
18
  - `Canopy-19` is a 0.2-second dead clip, and `Graminoid-03` has only six seconds of IR; both are empty because sampling starts near the 10th second.
@@ -23,6 +23,6 @@
23
 
24
  Use `scripts/make_random_splits.py` to hold out one burn site as the test set and divide the other site into training and validation sets.
25
 
26
- The script writes `Data/splits.csv` with one row per sampled frame and the columns `frame` and `split`. Split values are `train`, `val`, or `test`.
27
 
28
  The related clips documented in `markdown/flights.md` are treated as one inseparable flight block, so they cannot be divided between training and validation. The script randomly selects, using seed `42` by default, from all complete-block combinations that place 10–20% of the non-test frames in validation.
 
12
  └── homography/{homography.json,points.json}
13
  ```
14
 
15
+ `fXXXXX` is the original zero-based RGB frame index. RGB images are 2688×1512 and IR images are 640×360. Each mask is a flattened 2688×1512 row-major array of class labels `1`–`5`. `samples.csv` lists the sampled pairs. `splits_graminoid.csv` (Graminoid held out as the test set) and `splits_canopy.csv` (Canopy held out) give each frame's `train`, `val`, or `test` assignment in the two evaluation settings, and `annotation_label_by_split_graminoid.csv` / `annotation_label_by_split_canopy.csv` the class pixel shares per split.
16
 
17
  ## Known gaps
18
  - `Canopy-19` is a 0.2-second dead clip, and `Graminoid-03` has only six seconds of IR; both are empty because sampling starts near the 10th second.
 
23
 
24
  Use `scripts/make_random_splits.py` to hold out one burn site as the test set and divide the other site into training and validation sets.
25
 
26
+ The script writes `Data/splits.csv` with one row per sampled frame and the columns `frame` and `split`. Split values are `train`, `val`, or `test`. The released `splits_graminoid.csv` and `splits_canopy.csv` are that output for `--test Graminoid` and `--test Canopy`, renamed.
27
 
28
  The related clips documented in `markdown/flights.md` are treated as one inseparable flight block, so they cannot be divided between training and validation. The script randomly selects, using seed `42` by default, from all complete-block combinations that place 10–20% of the non-test frames in validation.
PseudoLabeling/GENERATED_LABELS.md CHANGED
@@ -1,17 +1,42 @@
1
- Automatically generated labels
2
 
3
- Everything in this PseudoLabeling/ folder relates to automatically generated labels. PseudoDataset10s/ and GeneratedAnnotations10s/ hold labels produced with the pseudo-labeling pipeline described in the paper, using the tool in annotation_tool/. Both folders follow the same layout and mask format as the hand-labeled Data/ folder in the repository root. The labels are generated rather than hand-annotated, so spot-check mask_overlay/ before relying on them.
4
 
5
- PseudoDataset10s (2,214 frames, 33 clips) covers frames that are not in Data/ at all. Three further frames were extracted from the source videos inside each 300-frame gap between two consecutive hand-labeled frames, at 50, 150 and 250 frames past the earlier one. It is meant as extra training data. Frames where the drone was moving between those two hand-labeled frames are left out. GeneratedAnnotations10s (775 frames, 31 clips) labels the frames already in Data/. Those frames are annotated by hand as well, so it is for reference rather than training.
6
 
7
- pseudo_samples.csv lists one row per new frame, with a split column to filter on before training.
 
 
 
8
 
9
- Entries for the ## Contents list.
10
 
11
- PseudoDataset10s/ holds generated labels for new frames sampled between the 10-second samples.
12
 
13
- GeneratedAnnotations10s/ holds generated labels for the frames already in Data/, for reference rather than training.
14
 
15
- annotation_tool/ holds the labeling pipeline that produced both. See annotation_tool/README.md.
16
 
17
- generate_pseudo_dataset.py is the script that produced PseudoDataset10s. Running it needs the source videos, which are not part of this release.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Automatically generated labels
2
 
3
+ 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/`.
4
 
5
+ 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.
6
 
7
+ | Folder | What it contains | Use it for | Frames |
8
+ |---|---|---|---|
9
+ | `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 |
10
+ | `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 |
11
 
12
+ ## Which frames the pseudo dataset adds
13
 
14
+ `../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.
15
 
16
+ 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.
17
 
18
+ `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.
19
 
20
+ ## Layout
21
+
22
+ 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.
23
+
24
+ - `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/`.
25
+
26
+ - `mask_overlay/`, `mask_color/`: the labels drawn over the RGB frame, and as a flat color image.
27
+
28
+ - `rgb/`, `ir/`, `ir_gray/`: the frame pair, with the de-colorized IR following the same convention as `../Data/`.
29
+
30
+ - `homography/`: the IR-to-RGB registration for the frame.
31
+
32
+ Classes are 1 = red, 2 = orange, 3 = yellow, 4 = green, 5 = black. Every pixel gets one of these five; there is no "unlabeled" value.
33
+
34
+ ## Contents of this folder
35
+
36
+ - `PseudoDataset10s/`: automatically generated labels for **new** frames sampled between the 10-second samples.
37
+
38
+ - `GeneratedAnnotations10s/`: automatically generated labels for the frames **already** in `../Data/`, for reference rather than training.
39
+
40
+ - `annotation_tool/`: the labeling pipeline that produced both; see `annotation_tool/README.md`.
41
+
42
+ - `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.
PseudoLabeling/annotation_tool/README.md CHANGED
@@ -1,7 +1,7 @@
1
  # Automatic RGB+IR Labeling Tool
2
 
3
  Generates per-pixel label annotations for registered RGB/IR wildfire footage
4
- (FilteredDataset10s format). A trained YOLO detector and IR hotspot thresholding propose
5
  candidate regions, and SAM turns those proposals into per-pixel masks. You give
6
  it a dataset folder; for each RGB/IR frame pair it writes a label mask in the same format
7
  as the dataset's own `mask/` files, plus overlay images.
@@ -21,8 +21,8 @@ as the dataset's own `mask/` files, plus overlay images.
21
 
22
  2. **Your data** (not included), in the layout described under
23
  "Running on your own data" below. For the original dataset, point
24
- at the folder that directly contains `Canopy/` and `Graminoid/`. Note it unzips
25
- double-nested, so that is usually `FilteredDataset10s/FilteredDataset10s`.
26
 
27
  3. **Python 3.10+**:
28
 
@@ -49,7 +49,7 @@ Everything in the dataset's train + test splits (775 pairs, about 3 hours on a G
49
  hours on CPU):
50
 
51
  ```bash
52
- python auto_annotate.py --dataset path/to/FilteredDataset10s/FilteredDataset10s --split train test --out MyAnnotations
53
  ```
54
 
55
  One clip (`--clip` is repeatable), or specific frames of it:
@@ -68,7 +68,8 @@ python auto_annotate.py --dataset path/to/dataset --all --out MyAnnotations
68
  Other options:
69
 
70
  - `--split` accepts `train`, `val`, `test` (any combination). It reads the frame list
71
- from `samples_splits.csv` at the dataset root, so it only works on the original dataset.
 
72
  - Interrupted run? Just run the same command again; finished frames are skipped. Use
73
  `--overwrite` to redo them.
74
  - `--no-copy-inputs` skips copying the rgb/ir frames into the output. This saves about
@@ -90,8 +91,8 @@ MyAnnotations/<Unit>/<Clip>/
90
  mask_overlay/fNNNNN.jpg labels blended over the RGB frame, for eyeballing
91
  mask_color/fNNNNN.png labels as a flat color image
92
  rgb/, ir/, homography/ copies of the inputs, so each output clip folder loads
93
- standalone in the dataset's own viewer.html (that file is
94
- at the root of FilteredDataset10s, not included here)
95
  ```
96
 
97
  **Mask format:** exactly 2688 x 1512 = 4,064,256 ASCII digit bytes. One digit per pixel,
 
1
  # Automatic RGB+IR Labeling Tool
2
 
3
  Generates per-pixel label annotations for registered RGB/IR wildfire footage
4
+ (the layout of `../../Data/`, the hand-labeled data at the root of this repository). A trained YOLO detector and IR hotspot thresholding propose
5
  candidate regions, and SAM turns those proposals into per-pixel masks. You give
6
  it a dataset folder; for each RGB/IR frame pair it writes a label mask in the same format
7
  as the dataset's own `mask/` files, plus overlay images.
 
21
 
22
  2. **Your data** (not included), in the layout described under
23
  "Running on your own data" below. For the original dataset, point
24
+ at the folder that directly contains `Canopy/` and `Graminoid/`, which in this
25
+ repository is `../../Data` relative to this folder.
26
 
27
  3. **Python 3.10+**:
28
 
 
49
  hours on CPU):
50
 
51
  ```bash
52
+ python auto_annotate.py --dataset ../../Data --split train test --out MyAnnotations
53
  ```
54
 
55
  One clip (`--clip` is repeatable), or specific frames of it:
 
68
  Other options:
69
 
70
  - `--split` accepts `train`, `val`, `test` (any combination). It reads the frame list
71
+ from `samples_splits.csv` at the dataset root, the annotators' working file, which is not
72
+ part of this release. On the released `../../Data/` use `--clip` or `--all` instead.
73
  - Interrupted run? Just run the same command again; finished frames are skipped. Use
74
  `--overwrite` to redo them.
75
  - `--no-copy-inputs` skips copying the rgb/ir frames into the output. This saves about
 
91
  mask_overlay/fNNNNN.jpg labels blended over the RGB frame, for eyeballing
92
  mask_color/fNNNNN.png labels as a flat color image
93
  rgb/, ir/, homography/ copies of the inputs, so each output clip folder loads
94
+ standalone in the dataset's own viewer.html
95
+ (`../../html/viewer.html` in this repository)
96
  ```
97
 
98
  **Mask format:** exactly 2688 x 1512 = 4,064,256 ASCII digit bytes. One digit per pixel,
PseudoLabeling/annotation_tool/auto_annotate.py CHANGED
@@ -1,6 +1,6 @@
1
  #!/usr/bin/env python3
2
  """
3
- Automatic labeling for registered RGB/IR frame pairs (FilteredDataset10s layout).
4
 
5
  Given a clip folder from the dataset —
6
  <Unit>/<Clip>/rgb/fNNNNN.png RGB frame (2688x1512)
@@ -43,7 +43,7 @@ Outputs, mirroring the dataset layout under --out:
43
  and loads in viewer.html
44
 
45
  Usage:
46
- python auto_annotate.py --dataset FilteredDataset10s/FilteredDataset10s \
47
  --clip Graminoid-08 --frames f00300
48
  python auto_annotate.py --dataset <root> --clip Canopy-10 (whole clip)
49
  python auto_annotate.py --dataset <root> --all (everything)
@@ -663,7 +663,7 @@ def main():
663
  raise SystemExit(
664
  f"Error: nothing to annotate under '{args.dataset}'. Expected <root>/<Unit>/<Clip>/rgb/ "
665
  "layout — pass the folder that directly contains Canopy/ and Graminoid/ "
666
- "(note: the dataset unpacks double-nested as FilteredDataset10s/FilteredDataset10s)."
667
  )
668
 
669
  print(f"PyTorch {torch.__version__} | CUDA: {torch.cuda.is_available()}")
 
1
  #!/usr/bin/env python3
2
  """
3
+ Automatic labeling for registered RGB/IR frame pairs (the layout of HotSight's Data/ folder).
4
 
5
  Given a clip folder from the dataset —
6
  <Unit>/<Clip>/rgb/fNNNNN.png RGB frame (2688x1512)
 
43
  and loads in viewer.html
44
 
45
  Usage:
46
+ python auto_annotate.py --dataset ../../Data \
47
  --clip Graminoid-08 --frames f00300
48
  python auto_annotate.py --dataset <root> --clip Canopy-10 (whole clip)
49
  python auto_annotate.py --dataset <root> --all (everything)
 
663
  raise SystemExit(
664
  f"Error: nothing to annotate under '{args.dataset}'. Expected <root>/<Unit>/<Clip>/rgb/ "
665
  "layout — pass the folder that directly contains Canopy/ and Graminoid/ "
666
+ "(in this repository: ../../Data relative to annotation_tool/)."
667
  )
668
 
669
  print(f"PyTorch {torch.__version__} | CUDA: {torch.cuda.is_available()}")
PseudoLabeling/generate_pseudo_dataset.py CHANGED
@@ -1,20 +1,23 @@
1
  #!/usr/bin/env python3
2
  """
3
- Pseudo-dataset generator for FilteredDataset10s.
4
 
5
- Samples NEW intermediate frames (frames that are not part of FilteredDataset10s)
6
  from the original RawVideos videos, builds a pseudo-dataset that
7
- follows the FilteredDataset10s conventions exactly, and (by default) runs the
8
  annotation tool (annotation_tool/auto_annotate.py) on the new frames.
9
 
10
  The actual frame extraction and RGB/IR pairing is delegated to the dataset
11
- team's canonical sampler, scripts/sample_dataset.py (RGB seek + sequential IR
12
  decode, IR partner = nearest by container timestamp to rgb_time + ir_offset —
13
  the same tool/convention that produced FilteredDataset10s; the old SRT-based
14
- sync is retired). Overlays are produced by scripts/make_mask_overlays.py
15
  (written as mask/fNNNNN_overlay.jpg, their convention). The scripts folder is
16
- located automatically next to this file's folder or inside it (--scripts-dir
17
- overrides).
 
 
 
18
 
19
  How a new frame is built
20
  ------------------------
@@ -48,7 +51,7 @@ Exclusions
48
  (Canopy-01, Canopy-02, Canopy-16, Canopy-19, ...).
49
  * Frames are only sampled inside exactly-300 anchor gaps, so any region the
50
  dataset filtered out (larger gaps) is skipped automatically.
51
- * The problematic video ranges documented in FilteredDataset10s/notes.md
52
  (IR "blend of IR and RGB" ranges and junk RGB ranges) are excluded
53
  explicitly as well, with inclusive endpoints.
54
 
@@ -107,7 +110,10 @@ import numpy as np
107
 
108
  SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
109
 
110
- DEFAULT_DATASET = os.path.join(SCRIPT_DIR, "FilteredDataset10s", "FilteredDataset10s")
 
 
 
111
  DEFAULT_VIDEOS = os.path.join(SCRIPT_DIR, "RawVideos")
112
  DEFAULT_OUT = os.path.join(SCRIPT_DIR, "PseudoDataset10s")
113
  ANNOTATE_SCRIPT = os.path.join(SCRIPT_DIR, "annotation_tool", "auto_annotate.py")
@@ -551,7 +557,7 @@ def main():
551
  "the RawVideos videos, with interpolated "
552
  "homographies, following FilteredDataset10s conventions.")
553
  p.add_argument("--dataset", default=DEFAULT_DATASET,
554
- help="FilteredDataset10s root (folder containing Canopy/ and Graminoid/).")
555
  p.add_argument("--videos", default=DEFAULT_VIDEOS,
556
  help="RawVideos root (passed to sample_dataset.py as --raw-root; "
557
  "must contain 'Livingston Place 18 19 20 March 2026/...').")
 
1
  #!/usr/bin/env python3
2
  """
3
+ Pseudo-dataset generator for HotSight (the hand-labeled data in ../Data/).
4
 
5
+ Samples NEW intermediate frames (frames that are not part of ../Data/)
6
  from the original RawVideos videos, builds a pseudo-dataset that
7
+ follows the ../Data/ conventions exactly, and (by default) runs the
8
  annotation tool (annotation_tool/auto_annotate.py) on the new frames.
9
 
10
  The actual frame extraction and RGB/IR pairing is delegated to the dataset
11
+ team's canonical sampler, ../scripts/sample_dataset.py (RGB seek + sequential IR
12
  decode, IR partner = nearest by container timestamp to rgb_time + ir_offset —
13
  the same tool/convention that produced FilteredDataset10s; the old SRT-based
14
+ sync is retired). Overlays are produced by ../scripts/make_mask_overlays.py
15
  (written as mask/fNNNNN_overlay.jpg, their convention). The scripts folder is
16
+ located automatically one level up from this file's folder (../scripts/) or
17
+ inside it (--scripts-dir overrides).
18
+
19
+ This script needs the source videos and the annotators' samples_splits.csv,
20
+ neither of which is part of the public release.
21
 
22
  How a new frame is built
23
  ------------------------
 
51
  (Canopy-01, Canopy-02, Canopy-16, Canopy-19, ...).
52
  * Frames are only sampled inside exactly-300 anchor gaps, so any region the
53
  dataset filtered out (larger gaps) is skipped automatically.
54
+ * The problematic video ranges documented in ../markdown/junk.md and ../markdown/blend.md
55
  (IR "blend of IR and RGB" ranges and junk RGB ranges) are excluded
56
  explicitly as well, with inclusive endpoints.
57
 
 
110
 
111
  SCRIPT_DIR = os.path.dirname(os.path.abspath(__file__))
112
 
113
+ # Hand-labeled data: ../Data in the released layout, FilteredDataset10s/FilteredDataset10s during development.
114
+ DEFAULT_DATASET = next((c for c in [os.path.join(os.path.dirname(SCRIPT_DIR), "Data"),
115
+ os.path.join(SCRIPT_DIR, "FilteredDataset10s", "FilteredDataset10s")]
116
+ if os.path.isdir(c)), os.path.join(os.path.dirname(SCRIPT_DIR), "Data"))
117
  DEFAULT_VIDEOS = os.path.join(SCRIPT_DIR, "RawVideos")
118
  DEFAULT_OUT = os.path.join(SCRIPT_DIR, "PseudoDataset10s")
119
  ANNOTATE_SCRIPT = os.path.join(SCRIPT_DIR, "annotation_tool", "auto_annotate.py")
 
557
  "the RawVideos videos, with interpolated "
558
  "homographies, following FilteredDataset10s conventions.")
559
  p.add_argument("--dataset", default=DEFAULT_DATASET,
560
+ help="Hand-labeled data root, the folder containing Canopy/ and Graminoid/ (default: ../Data next to this folder).")
561
  p.add_argument("--videos", default=DEFAULT_VIDEOS,
562
  help="RawVideos root (passed to sample_dataset.py as --raw-root; "
563
  "must contain 'Livingston Place 18 19 20 March 2026/...').")
scripts/README.md CHANGED
@@ -4,7 +4,7 @@ Run these from the root directory. Use `python scripts/<name>.py --help` to see
4
 
5
  - `sample_dataset.py` — samples synchronized RGB and IR frames from the original videos and writes `Data/samples.csv`.
6
 
7
- - `make_random_splits.py` — holds out one burn site for testing and writes `Data/splits.csv`, randomly selecting a 10–20% validation while keeping related flight groups together.
8
 
9
  - `filter_junk.py` — removes sampled frame pairs that fall within the unusable video ranges listed in `markdown/junk.md`. It modifies the dataset in place.
10
 
@@ -16,4 +16,4 @@ Run these from the root directory. Use `python scripts/<name>.py --help` to see
16
 
17
  - `make_mask_overlays.py` — draws each mask over its RGB image for quick visual inspection.
18
 
19
- - `analyze_label_distribution.py` — calculates per-clip mask-label percentages in `annotation_label_by_clip.csv`, passing `--splits splits.csv` also writes `annotation_label_by_split.csv`.
 
4
 
5
  - `sample_dataset.py` — samples synchronized RGB and IR frames from the original videos and writes `Data/samples.csv`.
6
 
7
+ - `make_random_splits.py` — holds out one burn site for testing and writes `Data/splits.csv`, randomly selecting a 10–20% validation while keeping related flight groups together. The released `Data/splits_graminoid.csv` and `Data/splits_canopy.csv` are its output for `--test Graminoid` and `--test Canopy`.
8
 
9
  - `filter_junk.py` — removes sampled frame pairs that fall within the unusable video ranges listed in `markdown/junk.md`. It modifies the dataset in place.
10
 
 
16
 
17
  - `make_mask_overlays.py` — draws each mask over its RGB image for quick visual inspection.
18
 
19
+ - `analyze_label_distribution.py` — calculates per-clip mask-label percentages in `annotation_label_by_clip.csv`, passing `--splits splits_graminoid.csv` (or `splits_canopy.csv`) also writes `annotation_label_by_split.csv`, released as `annotation_label_by_split_graminoid.csv` and `annotation_label_by_split_canopy.csv`.