TackleNet

A Biomechanical Practice-Video Benchmark for Tackle-Safety Assessment in American Football

This repository is the companion code release for the TackleNet benchmark dataset, accompanying the NeurIPS 2026 Datasets and Benchmarks submission.

Run all commands from the repository root. Scripts use package-relative imports (from eval.metrics import ...) that require the repo root to be on the Python path. Running from a subdirectory will produce ModuleNotFoundError.

Full dataset (~25 GB): https://www.kaggle.com/datasets/ahsanzaidi786/tacklenet
30-clip sample (~1 GB): https://www.kaggle.com/datasets/ahsanzaidi786/tacklenet-sample

---

Dataset at a Glance

Property	Value
Total clips	737 de-identified MP4 clips
Risky / Safe	259 risky (35.1%) / 478 safe (64.9%)
Practice sessions	30 sessions, 3 collection sites (PA, KS, FL)
Benchmark label	satt3_risk_binary derived from SATT3 component score
FPOC label	fpoc_frame — 0-indexed frame of First Point of Contact
Train / Val / Test	515 / 111 / 111 (stratified clip-level split, seed 42)
De-identification	RetinaFace + SAM 3.1 (face regions)
Data license	CC BY 4.0
Code license	MIT

---

Benchmark Tasks

Task	Goal	Primary Metric
Task A — Risk Classification	Predict tackle safety (risky/safe) from video clip	Risky recall
Task B — Contact Onset Localization	Predict FPOC frame index	Acc@±20 frames
Task C — Contact Window Analysis	Oracle analysis: GT FPOC window vs. full clip	PR-AUC delta

Task A and Task C use the frozen train/val/test split (splits/splits_frozen.csv). Task B methods are zero-shot and are evaluated on all 737 clips.

---

Repository Structure

tacklenet/
├── README.md
├── LICENSE
├── CITATION.cff
├── requirements.txt
├── CONTRIBUTING.md
├── .gitignore
├── data/
│   ├── __init__.py
│   └── loader.py                    # PyTorch Dataset for TackleNet clips
├── splits/
│   ├── annotations_satt_fpoc.csv    # Frozen annotation table (737 clips)
│   ├── splits_frozen.csv            # Official train/val/test split
│   ├── split_policy.md              # Split rationale and limitations
│   └── tacklenet_create_splits_v2.py
├── baselines/
│   ├── majority_baseline.py         # Task A: majority-class baseline
│   └── center_frame_baseline.py     # Task B: center-frame FPOC baseline
├── eval/
│   ├── metrics.py                   # Shared metric utilities
│   ├── evaluate_task_a.py           # Task A evaluator
│   └── evaluate_task_b.py           # Task B evaluator
├── examples/
│   ├── task_a_predictions_example.csv   # 10-row format demo — not full eval
│   └── task_b_predictions_example.csv   # 10-row format demo — not full eval
└── notebooks/
    └── tacklenet_usage.ipynb        # Training-free usage notebook

---

Quick Start

Interactive notebook (no setup required): https://www.kaggle.com/code/ahsanzaidi786/tacklenet-usage

1. Install dependencies

pip install -r requirements.txt

2. Download the dataset

# Full dataset (~25 GB)
kaggle datasets download ahsanzaidi786/tacklenet --unzip -p ./tacklenet_data

# 30-clip sample (~1 GB) — for pipeline verification only
kaggle datasets download ahsanzaidi786/tacklenet-sample --unzip -p ./tacklenet_sample

---

Task A — Risk Classification

Run the majority baseline

python baselines/majority_baseline.py \
    --annot   splits/annotations_satt_fpoc.csv \
    --splits  splits/splits_frozen.csv \
    --split   test \
    --out_dir outputs/majority_baseline

Outputs: majority_predictions.csv, majority_metrics.json

Evaluate Task A predictions

python eval/evaluate_task_a.py \
    --predictions outputs/majority_baseline/majority_predictions.csv \
    --splits      splits/splits_frozen.csv \
    --split       test \
    --out_dir     outputs/majority_eval

Outputs: task_a_metrics.json, task_a_classification_report.txt, task_a_confusion_matrix.png

Prediction file format

One row per test clip:

Column	Type	Description
clip_id	int	Clip identifier matching splits_frozen.csv
prob_risky	float [0,1]	Predicted risky-class probability
pred_label	int 0/1	Predicted binary label

See examples/task_a_predictions_example.csv for a 10-row format demo.

---

Task B — Contact Onset Localization

Task B methods are zero-shot and are evaluated on all 737 clips. The GRAZE pipeline (reference Task B method) is maintained separately:

https://github.com/AhsanZaidi12/GRAZE

GRAZE predictions can be evaluated directly with eval/evaluate_task_b.py.

Run the center-frame baseline

python baselines/center_frame_baseline.py \
    --annot     splits/annotations_satt_fpoc.csv \
    --video_dir path/to/tacklenet_data/videos \
    --out_dir   outputs/center_frame

Outputs: center_frame_predictions.csv (for the evaluator), center_frame_results.csv (per-clip detail)

Evaluate Task B predictions

python eval/evaluate_task_b.py \
    --predictions outputs/center_frame/center_frame_predictions.csv \
    --annot       splits/annotations_satt_fpoc.csv \
    --splits      splits/splits_frozen.csv \
    --split       all \
    --out_dir     outputs/center_frame_eval

Outputs: task_b_metrics.json, task_b_error_distribution.png

Smoke-test the evaluator format (no videos needed)

The example file contains only 10 rows and is intended for format validation only — not for benchmark reporting. Running against --split test (111 clips) will correctly show partial coverage.

python eval/evaluate_task_b.py \
    --predictions examples/task_b_predictions_example.csv \
    --annot       splits/annotations_satt_fpoc.csv \
    --splits      splits/splits_frozen.csv \
    --split       test \
    --out_dir     /tmp/test_eval_b

Expected output: n_covered=10, n_total=111, coverage=9.0%. This confirms the evaluator runs correctly; the low coverage is expected because the example file contains only 10 rows.

Prediction file format

One row per clip:

Column	Type	Description
clip_id	int	Clip identifier
pred_fpoc_frame	int	Predicted 0-indexed FPOC frame

See examples/task_b_predictions_example.csv for a 10-row format demo.

---

Task C — Contact Window Analysis

Task C is an oracle analysis comparing classifiers trained and evaluated on full clips against classifiers trained and evaluated on the GT FPOC 32-frame window (23 pre-contact + FPOC frame + 8 post-contact).

It uses the same Task A evaluation protocol. Train a classifier under each input condition, save its test predictions, and run evaluate_task_a.py on each. Compare the PR-AUC values.

---

Split Policy

Always use splits/splits_frozen.csv. Do not regenerate or substitute a custom split for benchmark-comparable results. See splits/split_policy.md for full details.

The split is stratified at the clip level (seed 42, 70/15/15). It is not session-grouped because reliable session identifiers were not available at split construction time. Session_ID is provided in the annotation file for transparency and post-hoc analysis only.

Verify the frozen split

python splits/tacklenet_create_splits_v2.py \
    --input  splits/annotations_satt_fpoc.csv \
    --output splits_check.csv \
    --seed   42

splits_check.csv must match splits/splits_frozen.csv exactly.

---

Label Convention

A clip is labeled risky (satt3_risk_binary = 1) if its SATT3 component score is 0 or 1, and safe (satt3_risk_binary = 0) if the score is 2 or 3.

---

FPOC Window Convention

All FPOC-windowed models use the locked asymmetric 32-frame window:

• 23 frames before FPOC + FPOC frame + 8 frames after = 32 frames total
• FPOC is placed at window index 23 (pre-contact heavy by design)
• Boundary handling: repeat-first / repeat-last frame clamping (never zero-pad)

Implemented in data/loader.py and demonstrated in notebooks/tacklenet_usage.ipynb.

---