# Combat action analyzers: source and protocol comparison

Checked September 15, 2026. This is a source-inspected selection of relevant projects, not an exhaustive inventory or a reproduced leaderboard. The exact repository recalled by the user remains unconfirmed. **BoxMind/BoxingWeb appears in the earlier local research**; UFC Fight Analyzer, ACM40960/Boxing, boxe.ml and NeuroCombat are other plausible matches.

Our comparison target is the compact pose-based event spotter: named joint filters, temporal layers, per-class event peaks and widths, followed by shared decoding. Its [historical study](/streaming-action-recognition-limits.md) measured labeled-window recall and background false alarms. It did not establish fully causal end-to-end operation.

## Source-inspected implementations

### BoxMind / BoxingWeb

The paper's TCN predicts hand-specific events and start/end offsets, then suppresses duplicate detections. **Detection F1 is .783 at temporal IoU .5**. Its pose/RGB attribute pipeline improves average attribute F1 from .681 with pose alone to .700 with guidance and expert heads. These are distinct metrics. The paper explicitly describes offline use. [Appendix B.2–B.3 and §4.4](https://arxiv.org/html/2601.11492v1)

The released loader instead consumes supplied `frame_begin`/`frame_end` annotations. The inspected repository provides attribute classification, not the paper's complete boundary-detection pipeline. [Event loading](https://github.com/gouba2333/BoxingWeb/blob/e06add251e31fe7db8db9fffcb705636484eb264/dataset/dataset_youtube.py#L271), [cropping](https://github.com/gouba2333/BoxingWeb/blob/e06add251e31fe7db8db9fffcb705636484eb264/dataset/dataset_youtube.py#L337), [fusion model](https://github.com/gouba2333/BoxingWeb/blob/e06add251e31fe7db8db9fffcb705636484eb264/model/mfusion.py#L8)

**Lesson:** boundary heads are closely related to ours; richer visual evidence and separate attributes are transferable hypotheses. The published system is not evidence that everyone else receives action cuts.

### ACM40960/Boxing

YOLO poses feed a **25-frame trailing buffer**, eight joints' normalized XY coordinates and an ordinary LSTM. Separate models classify punches and blocks. A wrist-motion gate, confidence threshold .7 and **60-frame recovery** decide when to output. The punch head has no background class; the block head has `negative`. Opponent contact is approximated through 2D geometry. [Architecture](https://github.com/ACM40960/Boxing/blob/c9d94ffa0ae24b8065c6c40f5f9f8dc453619676/app/scripts/two_player_analyzer.py#L11), [gating](https://github.com/ACM40960/Boxing/blob/c9d94ffa0ae24b8065c6c40f5f9f8dc453619676/app/scripts/two_player_analyzer.py#L262), [recovery](https://github.com/ACM40960/Boxing/blob/c9d94ffa0ae24b8065c6c40f5f9f8dc453619676/app/scripts/two_player_analyzer.py#L325)

The training notebook splits samples randomly 80/20 and selects the best validation checkpoint. Its reported validation scores do not supply held-out-session event recall, false alarms or boundary error. [Training notebook](https://github.com/ACM40960/Boxing/blob/c9d94ffa0ae24b8065c6c40f5f9f8dc453619676/main/script.ipynb)

**Lesson:** past-window inference is feasible, but classification still needs event handling. A long recovery period can suppress rapid combinations; 60 frames is frame-rate-dependent, not a fixed delay in milliseconds.

### imsoo/fight_detection

OpenPose and SORT maintain **32 frames per person**. Joint angles, angle changes and displacements feed two 34-unit LSTMs and a four-way classifier: standing, walking, punching, kicking. Training recordings were explicitly cut into actions before overlapping samples were generated. A geometric collision rule decides whether motion hits another person's box. [Data preparation](https://github.com/imsoo/fight_detection), [model](https://github.com/imsoo/fight_detection/blob/f6a435ce03bda03d1f722058c1fce44b231da254/server/action_train.py#L9), [collision rule](https://github.com/imsoo/fight_detection/blob/f6a435ce03bda03d1f722058c1fce44b231da254/server/src/people.cpp#L147)

**Lesson:** a useful coarse-action and interaction baseline, not a demonstrated fine-grained strike segmenter. No comparable full-stream event metric was located in the inspected material.

### boxe.ml

YOLOv8/ByteTrack feeds **wrist-speed hysteresis segmentation**, then an LSTM classifies straight/hook/uppercut and geometry assigns lead/rear. This directly implements “cut movement, then classify.” The inspected path uses symmetric five-frame smoothing, completed segments and peak-adjacent samples. [Pipeline](https://github.com/leonardosm14/boxe.ml/blob/main/boxe.py#L128)

Its report gives **83% three-class accuracy and .78 macro F1 on 899 BoxingVI test samples**. These are classification metrics, not event detection over full recordings. Its shared window builder also matches training's ten observed frames plus padding to 25; a dense 25-frame inference input was a distribution mismatch. [Results](https://github.com/leonardosm14/boxe.ml/blob/main/docs/paper/sections/results.typ), [window builder](https://github.com/leonardosm14/boxe.ml/blob/main/utils/boxe_utils.py#L62)

**Lesson:** a concrete proposal-first alternative. Motion cuts still need testing against feints, rapid combinations and missed low-amplitude strikes. Its inspected implementation is offline.

### RTM-FASS

The Muay Thai project uses a **15-frame buffer**, linear frame encoders, a bidirectional LSTM and temporal attention. It predicts eight strike types plus hit/miss; a .65 threshold and ten-frame cooldown govern output. No learned boundary head or explicit background strike class appears in the inspected architecture. [Model](https://github.com/asaravanabavan/RTM-FASS/blob/main/src/strike_model.py#L16), [runtime](https://github.com/asaravanabavan/RTM-FASS/blob/main/src/strike_detector.py#L111)

Bidirectionality inside a completed trailing window does not inherently read after the current time; it would, however, use later frames for predictions assigned to earlier positions in that window. The documented weight-download link was a placeholder. No comparable event evaluation was located. [Model availability](https://github.com/asaravanabavan/RTM-FASS/blob/main/models/README.md)

### UFC Fight Analyzer

The name fits the recollection, but the inspected 122-line Python program performs YOLO person tracking and keypoint/skeleton display. It contains no action classifier or learned action segmentation. This is an input-processing prototype rather than a measured recognition competitor. [Complete implementation](https://github.com/dscueva/ufc_fight_analyzer/blob/main/ufc_fight_analyzer/analyze_fight.py)

## Related systems with different evidence

- **TAPE:** its author describes proposals, classification and human verification, with a few thousand verdicts and nested leave-one-fight-out evaluation. No public source repository was linked in the inspected writeup. Treat its metrics as author reports, not an independently reproduced baseline. Its review-assisted design is relevant to building reliable training data and traceable statistics. [Author's technical writeup](https://www.juyoungkim.net/posts/tape-mma-fight-intelligence)
- **NeuroCombat:** the author's link resolves to `AdityaK404/NeuroCombat-V_9`, which returned 404 during this review. Architecture and performance could not be audited. [Author profile containing the post](https://in.linkedin.com/in/sana-naaz-0a2723333), [referenced repository](https://github.com/AdityaK404/NeuroCombat-V_9)
- **BoxComm, April 2026 preprint:** describes causal-TCN punch detection followed by VideoMAEv2/TCN classification, then commentary generation. Detection training mixes public and proprietary data. Its reported commentary metrics do not provide a matched detector accuracy/false-alarm/latency comparison. This is a concrete published example of modern features combined with a streaming detector. [§5.1 and experiments](https://arxiv.org/html/2604.04419v1#S5.SS1)
- **MMAction2:** a Python toolbox supporting recognition and localization, not a pretrained MMA strike analyzer. Its name is another possible source of the recollection. [Official overview](https://mmaction2.readthedocs.io/en/latest/get_started/overview.html)

## What would make a fair comparison?

Use the same unseen fights and actor tracks, explicit event definitions, independently annotated timing and background, and the same allowed lookahead. Measure event recall, duplicate/missed events, false alarms per minute, boundary error and actual emission delay. Then compare a learned event head, motion-proposal-plus-classifier pipeline, and rolling-window classifier with its gates. Published clip accuracy, pose-tracking demos, commentary scores and our current window-hit recall do not form a shared leaderboard.
