Create a baseline object detection pipeline

training/object-detection/.dvc/.gitignore

@@ -0,0 +1,3 @@
+/config.local
+/tmp
+/cache

training/object-detection/.dvc/config

@@ -0,0 +1,4 @@
+[core]
+    remote = storage
+['remote "storage"']
+    url = s3://salmonvision-dvc/rgb_object_detection

training/object-detection/.dvcignore

@@ -0,0 +1,3 @@
+# Add patterns of files dvc should ignore, which could improve
+# the performance. Learn more at
+# https://dvc.org/doc/user-guide/dvcignore

training/object-detection/.python-version

@@ -0,0 +1 @@
+3.8

training/object-detection/README.md

@@ -0,0 +1,35 @@
+# Object Detection
+
+Training pipeline to train the SalmonVision object detection model.
+
+Install uv:
+```bash
+curl -LsSf https://astral.sh/uv/install.sh | sh
+```
+
+Install DVC:
+```bash
+uv tool install dvc
+```
+
+Install the module:
+```bash
+uv pip install -e .
+```
+
+Check dvc.yaml for the full pipeline.
+
+Run the following to run specific stages of the pipeline:
+```bash
+dvc repro stage_name
+```
+
+For example, building the model input annotations:
+```bash
+dvc repro build_model_input
+```
+
+Run tests with
+```
+uv run pytest
+```

training/object-detection/config/salmon_yolo.yaml

@@ -0,0 +1,28 @@
+# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
+# Classes updated on 2026-02-12
+path: /training/export_combined_bear_kitwanga_yolo  # dataset root dir
+train: train.txt
+val: val.txt
+test: test.txt
+
+# Classes
+names:
+  0: Coho
+  1: Bull
+  2: Rainbow
+  3: Sockeye
+  4: Pink
+  5: Whitefish
+  6: Chinook
+  7: Shiner
+  8: Pikeminnow
+  9: Chum
+  10: Steelhead
+  11: Lamprey
+  12: Cutthroat
+  13: Stickleback
+  14: Sculpin
+  15: Jack_Coho
+  16: Jack_Chinook
+  17: Otter
+  18: Sucker

training/object-detection/data/01_raw/.gitignore

@@ -0,0 +1,4 @@
+/labelstudio_annos
+/salmon_vid_counts.csv
+/salmon_vid_counts_summary.csv
+/sv_water_conditions

training/object-detection/data/01_raw/salmon_vid_counts.csv.dvc

@@ -0,0 +1,5 @@
+outs:
+- md5: b6136eee30a358d7473a160197bc91be
+  size: 6465051
+  hash: md5
+  path: salmon_vid_counts.csv

training/object-detection/data/01_raw/salmon_vid_counts_summary.csv.dvc

@@ -0,0 +1,5 @@
+outs:
+- md5: c65c2ba8214d26905ccb9608e9071b93
+  size: 1031
+  hash: md5
+  path: salmon_vid_counts_summary.csv

training/object-detection/data/01_raw/sv_water_conditions.dvc

@@ -0,0 +1,6 @@
+outs:
+- md5: 717ade5c2fd57a763609ab55d66bf351.dir
+  size: 199437
+  nfiles: 10
+  hash: md5
+  path: sv_water_conditions

training/object-detection/data/02_interim/.gitignore

@@ -0,0 +1,3 @@
+/yolo_annos
+/yolo_annos_unpacked
+/yolo_condition_negatives

training/object-detection/data/03_processed/.gitignore

@@ -0,0 +1 @@
+/splits_baseline

training/object-detection/dvc.lock

@@ -0,0 +1,143 @@
+schema: '2.0'
+stages:
+  update_raw:
+    cmd: rclone sync  -P --filter "- /salm_dataset*/**" --filter "+ *.json" 
+      --filter "- *.zip" aws:salmonvision-ml-datasets/rgb/raw/ 
+      data/01_raw/labelstudio_annos
+    deps:
+    - path: s3://salmonvision-ml-datasets/rgb/raw
+      hash: md5
+      md5: 9aab20758f57f39971a38c9a676dad27.dir
+      size: 720238505
+      nfiles: 65
+    outs:
+    - path: data/01_raw/labelstudio_annos
+      hash: md5
+      md5: 188b71c31be326fe36d209cc52c23337.dir
+      size: 379352848
+      nfiles: 50
+  split_data:
+    cmd: rm -rf data/02_interim/yolo_annos_unpacked && scripts/unpack_annos.sh 
+      data/02_interim/yolo_annos data/02_interim/yolo_annos_unpacked && 
+      scripts/unpack_annos.sh data/02_interim/yolo_condition_negatives 
+      data/02_interim/yolo_annos_unpacked && scripts/make_splits.py 
+      --labels-root data/02_interim/yolo_annos_unpacked --out-dir 
+      data/03_processed/splits_baseline --sites tankeeah kitwanga bear --seed 42
+      --train-frac 0.8 --val-frac 0.1 --test-frac 0.1
+    deps:
+    - path: data/01_raw/salmon_vid_counts.csv
+      hash: md5
+      md5: b6136eee30a358d7473a160197bc91be
+      size: 6465051
+    - path: data/01_raw/salmon_vid_counts_summary.csv
+      hash: md5
+      md5: c65c2ba8214d26905ccb9608e9071b93
+      size: 1031
+    - path: data/02_interim/yolo_annos
+      hash: md5
+      md5: 88551957fa5e19f676175f29606a613a.dir
+      size: 308350474
+      nfiles: 5
+    - path: data/02_interim/yolo_condition_negatives
+      hash: md5
+      md5: c3505d079516f38874bdb42d2182cb46.dir
+      size: 338928
+      nfiles: 3
+    - path: scripts/make_splits.py
+      hash: md5
+      md5: 3921729fec8a5aa6c1eb25c54196a658
+      size: 650
+    - path: src/object_detection/splits
+      hash: md5
+      md5: 8ca8d1d4f2962a0e27f6fd50da90a8d2.dir
+      size: 35847
+      nfiles: 6
+    - path: src/object_detection/utils
+      hash: md5
+      md5: 5828afbf8a25f11a0f8e5bbc6c0065bf.dir
+      size: 762
+      nfiles: 4
+    outs:
+    - path: data/03_processed/splits_baseline
+      hash: md5
+      md5: e821f72b0bf42a09afafb688ce927ac4.dir
+      size: 16966114
+      nfiles: 5
+  build_model_input:
+    cmd: scripts/yolo_converter_ls_video.py data/01_raw/labelstudio_annos 
+      --data-yaml config/salmon_yolo.yaml --out data/02_interim/yolo_annos_fs 
+      --out-shards data/02_interim/yolo_annos --empty-list 
+      data/02_interim/yolo_annos/empty_vids.txt --shard-size 100000 --pattern 
+      '**/*.json' --include-sites tankeeah kitwanga bear --frame-stride 3 
+      --frame-offset-mode video_hash --include-negatives --negative-ratio 0.10 
+      --negatives-per-video 11
+    deps:
+    - path: config/salmon_yolo.yaml
+      hash: md5
+      md5: f453f5dc54f1743eaedcc3ab117d269e
+      size: 547
+    - path: data/01_raw/labelstudio_annos
+      hash: md5
+      md5: 188b71c31be326fe36d209cc52c23337.dir
+      size: 379352848
+      nfiles: 50
+    - path: scripts/yolo_converter_ls_video.py
+      hash: md5
+      md5: 95696514ec77c20c6491c5af2ccb46a5
+      size: 118
+    - path: src/object_detection/utils
+      hash: md5
+      md5: 5828afbf8a25f11a0f8e5bbc6c0065bf.dir
+      size: 762
+      nfiles: 4
+    - path: src/object_detection/yolo_ls
+      hash: md5
+      md5: 80ca7160d1912861bbc2a013a45ac5d5.dir
+      size: 52757
+      nfiles: 10
+    outs:
+    - path: data/02_interim/yolo_annos
+      hash: md5
+      md5: 88551957fa5e19f676175f29606a613a.dir
+      size: 308350474
+      nfiles: 5
+  unpack_annos:
+    cmd: rm -r data/02_interim/yolo_annos_unpacked || scripts/unpack_annos.sh 
+      data/02_interim/yolo_annos data/02_interim/yolo_annos_unpacked
+    deps:
+    - path: data/02_interim/yolo_annos
+      hash: md5
+      md5: 473a8eedf5fd4cc5d5b8b6d1f80681e7.dir
+      size: 293713920
+      nfiles: 3
+  build_condition_negatives:
+    cmd: scripts/create_condition_negatives.py --conditions-csv 
+      data/01_raw/sv_water_conditions/SV_conditions_tracking_tankeeah_2025.csv 
+      data/01_raw/sv_water_conditions/SV_conditions_tracking_bear_2025.csv 
+      data/01_raw/sv_water_conditions/SV_conditions_tracking_kitwanga_2025.csv 
+      --out-dir data/02_interim/yolo_condition_negatives --frames-per-video 20 
+      --frame-stride 3 --frame-offset-mode video_hash --shard-size 100000
+    deps:
+    - path: data/01_raw/sv_water_conditions
+      hash: md5
+      md5: 717ade5c2fd57a763609ab55d66bf351.dir
+      size: 199437
+      nfiles: 10
+    - path: src/object_detection/negatives/cli.py
+      hash: md5
+      md5: d04fcc8b316529ebefeec2830b288197
+      size: 2142
+    - path: src/object_detection/negatives/conditions.py
+      hash: md5
+      md5: f037be2028fead4a5a9e88f95799254a
+      size: 18950
+    - path: src/object_detection/yolo_ls/shards.py
+      hash: md5
+      md5: 9a1e97420b95c978e10ec236c2645337
+      size: 1266
+    outs:
+    - path: data/02_interim/yolo_condition_negatives
+      hash: md5
+      md5: c3505d079516f38874bdb42d2182cb46.dir
+      size: 338928
+      nfiles: 3

training/object-detection/dvc.yaml

@@ -0,0 +1,142 @@
+stages:
+  update_raw:
+    cmd: >-
+      rclone sync 
+      -P
+      --filter "- /salm_dataset*/**"
+      --filter "+ *.json"
+      --filter "- *.zip"
+      aws:salmonvision-ml-datasets/rgb/raw/
+      data/01_raw/labelstudio_annos
+    deps:
+      - s3://salmonvision-ml-datasets/rgb/raw
+    outs:
+      - data/01_raw/labelstudio_annos
+    frozen: true
+  build_model_input:
+    cmd: >-
+      scripts/yolo_converter_ls_video.py
+      data/01_raw/labelstudio_annos
+      --data-yaml config/salmon_yolo.yaml
+      --out data/02_interim/yolo_annos_fs
+      --out-shards data/02_interim/yolo_annos
+      --empty-list data/02_interim/yolo_annos/empty_vids.txt
+      --shard-size 100000
+      --pattern '**/*.json'
+      --include-sites tankeeah kitwanga bear
+      --frame-stride 3
+      --frame-offset-mode video_hash
+      --include-negatives
+      --negative-ratio 0.10
+      --negatives-per-video 11
+    deps:
+      - scripts/yolo_converter_ls_video.py
+      - src/object_detection/yolo_ls
+      - src/object_detection/utils
+      - data/01_raw/labelstudio_annos
+      - config/salmon_yolo.yaml
+    outs:
+      - data/02_interim/yolo_annos
+  build_condition_negatives:
+    cmd: >-
+      scripts/create_condition_negatives.py
+      --conditions-csv data/01_raw/sv_water_conditions/SV_conditions_tracking_tankeeah_2025.csv
+      data/01_raw/sv_water_conditions/SV_conditions_tracking_bear_2025.csv
+      data/01_raw/sv_water_conditions/SV_conditions_tracking_kitwanga_2025.csv
+      --out-dir data/02_interim/yolo_condition_negatives
+      --frames-per-video 20
+      --frame-stride 3
+      --frame-offset-mode video_hash
+      --shard-size 100000
+    deps:
+      - src/object_detection/negatives/conditions.py
+      - src/object_detection/negatives/cli.py
+      - src/object_detection/yolo_ls/shards.py
+      - data/01_raw/sv_water_conditions
+    outs:
+      - data/02_interim/yolo_condition_negatives
+  split_data:
+    cmd: >-
+      rm -rf data/02_interim/yolo_annos_unpacked &&
+      scripts/unpack_annos.sh
+      data/02_interim/yolo_annos
+      data/02_interim/yolo_annos_unpacked &&
+      scripts/unpack_annos.sh
+      data/02_interim/yolo_condition_negatives
+      data/02_interim/yolo_annos_unpacked &&
+      scripts/make_splits.py
+      --labels-root data/02_interim/yolo_annos_unpacked
+      --out-dir data/03_processed/splits_baseline
+      --sites tankeeah kitwanga bear
+      --seed 42
+      --train-frac 0.8 --val-frac 0.1 --test-frac 0.1
+    deps:
+      - scripts/make_splits.py
+      - src/object_detection/splits
+      - src/object_detection/utils
+      - data/02_interim/yolo_annos
+      - data/02_interim/yolo_condition_negatives
+      - data/01_raw/salmon_vid_counts.csv
+      - data/01_raw/salmon_vid_counts_summary.csv
+    outs:
+      - data/03_processed/splits_baseline
+  build_video_metadata_index:
+    cmd: >-
+      scripts/build_video_metadata_index.py
+      --json-dir data/01_raw/labelstudio_annos
+      --out-csv data/02_interim/video_metadata_index.csv
+    deps:
+      - src/object_detection/metadata/index.py
+      - src/object_detection/metadata/cli.py
+      - data/01_raw/labelstudio_annos
+    outs:
+      - data/02_interim/video_metadata_index.csv
+  pack_split_dataset:
+    cmd: >-
+      scripts/pack_split_dataset.py
+      --splits-dir data/03_processed/splits_baseline
+      --labels-root data/02_interim/yolo_annos_unpacked
+      --shards-root ${config.drive}/salmon_dataset/dataset_sharded/shards
+      --manifests-root ${config.drive}/salmon_dataset/dataset_sharded/manifests
+      --temp-video-dir ${config.drive}/salmon_dataset/tmp_videos
+      --metadata-csv data/02_interim/video_metadata_index.csv
+      data/02_interim/yolo_condition_negatives/condition_negative_video_metadata.csv
+      --data-yaml config/salmon_yolo.yaml
+      --bucket prod-salmonvision-edge-assets-labelstudio-source
+      --image-ext .jpg
+      --manifest-csv data/03_processed/packed_dataset_manifest.csv
+      --splits train val test
+      --shard-size 100000
+    deps:
+      - src/object_detection/frames/parsing.py
+      - src/object_detection/frames/extractor.py
+      - src/object_detection/frames/cli.py
+      - src/object_detection/yolo_ls/shards.py
+      - data/03_processed/splits_baseline
+      - data/02_interim/yolo_annos_unpacked
+      - data/02_interim/video_metadata_index.csv
+      - data/02_interim/yolo_condition_negatives/condition_negative_video_metadata.csv
+      - config/salmon_yolo.yaml
+    params:
+      - config.drive
+    outs:
+      - ${config.drive}/salmon_dataset/dataset_sharded/shards
+      - ${config.drive}/salmon_dataset/dataset_sharded/manifests
+      - data/03_processed/packed_dataset_manifest.csv
+  unpack_split_dataset:
+    cmd: >-
+      rm -rf ${config.drive}/salmon_dataset/yolo_workdir &&
+      mkdir -p ${config.drive}/salmon_dataset/yolo_workdir &&
+      scripts/unpack_annos.sh
+      ${config.drive}/salmon_dataset/dataset_sharded/shards
+      ${config.drive}/salmon_dataset/yolo_workdir &&
+      cp ${config.drive}/salmon_dataset/dataset_sharded/manifests/train.txt ${config.drive}/salmon_dataset/yolo_workdir/train.txt &&
+      cp ${config.drive}/salmon_dataset/dataset_sharded/manifests/val.txt ${config.drive}/salmon_dataset/yolo_workdir/val.txt &&
+      cp ${config.drive}/salmon_dataset/dataset_sharded/manifests/test.txt ${config.drive}/salmon_dataset/yolo_workdir/test.txt &&
+      cp ${config.drive}/salmon_dataset/dataset_sharded/manifests/data.yaml ${config.drive}/salmon_dataset/yolo_workdir/data.yaml
+    deps:
+      - ${config.drive}/salmon_dataset/dataset_sharded/shards
+      - ${config.drive}/salmon_dataset/dataset_sharded/manifests
+    params:
+      - config.drive
+    frozen: true