FIFA + vailá Continuation Skill

Companion files:

• Maintenance rule: when you edit any *.py during this continuation, also update version/date metadata (global version from vaila.py), root README.md Last updated:, and matching help .md/.html. See AGENTS.md (“Mandatory: Update metadata on any script change”).

• ~/data/vaila/AGENT_HISTORY.md — what was done, current state, decisions.

• ~/data/vaila/HISTORY.md — chronological log of the session.

• .claude/skills/fifa-skeletal-tracking/SKILL.md — full FIFA pipeline reference.

• .claude/skills/sam3-video/SKILL.md — SAM 3 video segmentation reference.

• .claude/skills/soccer-field-keypoints-yolo/SKILL.md — pitch-keypoint training/inference.

• docs/fifa_workflow.md §4.5 — external merged unified/ dataset, QA export, dedupe, yolo pose train (weights → soccerfield_keypoints_ai).

• .claude/skills/yolo-fb-gui-cli/SKILL.md — YOLO + FB chooser + GUI→CLI mirror.

• docs/sessions/2026-07-06-yolo-fb-gui-cli-mirror.md — v0.3.72 session log.

• docs/sessions/2026-07-07-sapiens-output-dir-fix.md — v0.3.76 Sapiens2 single output dir.

Companion repo (the official starter kit): ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026.

---

Objective

Continue from a known-good baseline and execute the improvement path:

1. SAM 3 per-player video masks
2. Per-frame soccer pitch keypoints for camera refinement
3. Per-frame DLT export for rec2d.py / rec3d.py
4. (Optional) fifa baseline (SAM 3D Body) → fifa pack for Codabench

---

Known-good baseline (validated 2026-04-25)

• pyproject.toml (portable, one file for every OS) synced with the cuda dependency group (PyTorch cu129).
• uv sync --no-group cpu --group cuda --extra sam succeeded.
• vaila/models/sam3/sam3.pt and sam3.1_multiplex.pt present.
• cv2.VideoCapture is back after uv pip install --reinstall opencv-python==4.10.0.84.
• SAM 3 smoke test: 32 frames of ARG_CRO_000737.mp4 written to ~/data/FIFA/outputs_sam3_smoke/processed_sam_20260425_183925/.
• SAM 3 60-frame test: 37 person IDs tracked, sam_points.csv produced. Output: ~/data/FIFA/outputs_sam3_test1/processed_sam_20260425_213120/ARG_CRO_000737/.

Pitch-keypoint detector status (2026-04-25 21:42)

Model	Status	Path
Old football_pitch32_best.pt	COLLAPSED — all 32 kp in ~30 px cluster (mosaic-trained)	vaila/models/soccerfield_keypoints_yolo/football_pitch32_best.pt
New 50-ep recipe-A best.pt	Geometry recovered — kp span full image; pose_mAP50=0.17, box_mAP50=0.83	vaila/models/runs/pose_fifa/pitch32_recipeA_50ep/weights/best.pt
400-ep continuation	Training in background (PID 2070719)	vaila/models/runs/pose_fifa/pitch32_recipeA_400ep/weights/best.pt (when done)

Training recipe that escapes the collapse:

uv run yolo pose train \
  model=~/data/vaila/yolo26s-pose.pt \
  data=vaila/models/hf_datasets/football-pitch-detection/data/data.yaml \
  epochs=800 imgsz=1280 batch=8 \
  mosaic=0.0 mixup=0.0 close_mosaic=0 erasing=0.0 \
  pose=25.0 kobj=2.0 device=0 patience=80 \
  project=vaila/models/runs/pose_fifa name=pitch32_recipeA

Monitoring the running 400-ep job:

tail -1 ~/data/vaila/vaila/models/runs/pose_fifa/pitch32_recipeA_400ep/results.csv \
  | awk -F',' '{printf "epoch=%s box_mAP50=%s pose_mAP50=%s pose_mAP50-95=%s\n",$1,$11,$15,$16}'

# live progress
tail -f ~/data/vaila/vaila/models/runs/pose_fifa/train_recipeA_400ep.log

Sanity test for a freshly trained best.pt (CPU-friendly, no GPU):

CUDA_VISIBLE_DEVICES="" uv run python -m vaila.soccerfield_keypoints_ai \
  --mode video \
  -i ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data/videos/ARG_CRO_000737.mp4 \
  -o ~/data/FIFA/outputs_pitch_kps_v3 \
  --backend ultralytics \
  --weights ~/data/vaila/vaila/models/runs/pose_fifa/pitch32_recipeA_400ep/weights/best.pt \
  --imgsz 1280 --conf 0.20 --draw-min-conf 0.30 \
  --device cpu --stride 30 --max-frames 5 --overlay-video

Look at field_keypoints_video.csv and confirm kp x ranges across ~the whole image width (not collapsed).

---

Resume Sequence (paste in this order)

0. Enter the project

cd ~/data/vaila
source .venv/bin/activate     # optional — `uv run` works without it

1. Re-confirm the CUDA group + extras (idempotent)

bash bin/setup_pyproject.sh --target=cuda --yes
uv sync --no-group cpu --group cuda --extra sam

2. Make sure SAM 3 weights are still there

uv run vaila/vaila_sam.py --download-weights   # no-op if already present
ls vaila/models/sam3/sam3.pt

3. Quick smoke test (1 video, 32 frames) — only if a fresh sanity check is needed

uv run vaila/vaila_sam.py \
  -i ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data/videos/ARG_CRO_000737.mp4 \
  -o ~/data/FIFA/outputs_sam3_smoke \
  -t person --max-frames 32

4. Run SAM 3 batch over all FIFA videos

Default uses subprocess-per-video isolation (clean GPU between clips):

uv run vaila/vaila_sam.py \
  -i ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data/videos \
  -o ~/data/FIFA/outputs_sam3 \
  -t person \
  --max-frames 64 \
  --max-input-long-edge 1280 \
  --postprocess-points foot

Useful flags:

• --frame-by-frame — lower VRAM (CUDA only).
• --no-isolate-batch — same-process batch (debug only).
• -t "team in red", -t goalkeeper, etc. — text presets accepted by SAM 3.

5. Soccer-pitch keypoints (need EITHER Roboflow OR a YOLO best.pt)

5A. Roboflow backend (no local weights needed)

export ROBOFLOW_API_KEY="YOUR_KEY"
uv run python -m vaila.soccerfield_keypoints_ai \
  --mode video \
  -i ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data/videos/ARG_CRO_000737.mp4 \
  -o ~/data/FIFA/outputs_pitch_kps \
  --backend roboflow \
  --roboflow-model-id football-field-detection-f07vi/14 \
  --conf 0.3 --draw-min-conf 0.05 \
  --stride 1 --max-frames 300 \
  --overlay-video

5B. Local Ultralytics YOLO-pose

uv run python -m vaila.soccerfield_keypoints_ai \
  --mode video \
  -i ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data/videos/ARG_CRO_000737.mp4 \
  -o ~/data/FIFA/outputs_pitch_kps \
  --backend ultralytics \
  --weights /ABS/PATH/TO/best.pt \
  --imgsz 1280 --conf 0.3 --draw-min-conf 0.05 \
  --device 0 --stride 1 --overlay-video

To train a best.pt from the vendored dataset, follow .claude/skills/soccer-field-keypoints-yolo/SKILL.md (vaila/models/hf_datasets/football-pitch-detection/data/data.yaml, kpt_shape: [32, 3], imgsz=1280, mosaic=0, erasing=0).

6. Calibration (manual first, automatic later)

# List the semantic point names accepted by the calibrator
uv run python -m vaila.soccerfield_calib --list-keypoints

# Manual interactive calibration on a single FIFA clip
uv run python -m vaila.soccerfield_calib \
  -v ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data/videos/ARG_CRO_000737.mp4 \
  -o ~/data/FIFA/outputs_pitch_calib \
  --data-root ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data

The detector outputs generic p1, p2, …. Until a mapping (p_i → semantic name) is published, the safe path is manual calibration on a few key frames and reuse of the homography for the whole clip when the camera is roughly static.

7. Per-frame DLT for rec2d.py / rec3d.py

Once the FIFA cameras/*.npz is populated (either from the official dataset or from fifa baseline --export-camera):

uv run vaila/vaila_sam.py fifa dlt-export \
  --cameras-dir ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data/cameras \
  --output-dir   ~/data/FIFA/outputs_dlt_per_frame

Use the resulting per-frame .dlt2d / .dlt3d with vaila/rec2d.py and vaila/rec3d.py (broadcast / moving cameras). For genuinely fixed cameras, use rec2d_one_dlt2d.py / rec3d_one_dlt3d.py instead.

8. (Optional) Full FIFA pipeline + Codabench submission

# Once and only once (clones sam_3d_body + downloads gated weights)
uv run hf auth login                        # see workaround below if hf is broken
bash bin/setup_fifa_sam3d.sh
uv sync --no-group cpu --group cuda --extra sam --extra fifa

uv run vaila/vaila_sam.py fifa bootstrap \
  --videos-dir ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data/videos \
  --data-root  ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data
uv run vaila/vaila_sam.py fifa prepare    --data-root ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data
uv run vaila/vaila_sam.py fifa boxes      --data-root ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data
uv run vaila/vaila_sam.py fifa preprocess --data-root ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data
uv run vaila/vaila_sam.py fifa baseline   --data-root ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data \
                                          --output    ~/data/FIFA/outputs/submission_full.npz \
                                          --export-camera
uv run vaila/vaila_sam.py fifa dlt-export --cameras-dir ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data/cameras \
                                          --output-dir  ~/data/FIFA/outputs/dlt_per_frame
uv run vaila/vaila_sam.py fifa pack       --submission-full ~/data/FIFA/outputs/submission_full.npz \
                                          --data-root       ~/data/FIFA/FIFA-Skeletal-Tracking-Starter-Kit-2026/data \
                                          --output-dir      ~/data/FIFA/outputs \
                                          --split val

---

Recovery / known issues

• hf auth login fails inside .venv with ModuleNotFoundError: No module named 'typer'. Either set HF_TOKEN=... in .env, or run hf auth login from a separate pipx install huggingface_hub install.
• cv2.VideoCapture missing after a fresh sync: uv pip install --reinstall opencv-python==4.10.0.84.
• SAM 3 batch CUDA OOM cascade: already mitigated in vaila_sam.py via subprocess-per-video isolation. If you still OOM, add --frame-by-frame and lower --max-input-long-edge.

---

Next-step priorities (highest expected MPJPE win first)

1. Per-frame camera refinement with pitch keypoints (Section 5+6).
2. Cleaner player boxes/masks with SAM 3 batch (Section 4).
3. Only then touch fifa baseline and 3D refinement (Section 8).

---

Fallbacks when the retrain plateaus

1. Push pose loss higher (pose=40 kobj=4), reset from pitch32_recipeA_50ep/best.pt, 200–300 epochs.
2. Bigger backbone (yolo26m-pose.pt / yolo26x-pose.pt) with lr0=1e-4 amp=False to dodge NaN.
3. Keypoint-safe augmentation: degrees=10 translate=0.1 scale=0.5 — never mosaic or mixup.
4. Roboflow drop-in (no local training): --backend roboflow --roboflow-model-id football-field-detection-f07vi/14 with ROBOFLOW_API_KEY.
5. Manual homography: 4–6 keyframes/video with vaila/soccerfield_calib.py, cv2.findHomography, interpolate.
6. External SOTA as alternative pitch detector:
   • https://github.com/MM4SPA/PnLCalib (Apache-2)
   • https://github.com/mguti97/No-Bells-Just-Whistles (MIT)