Work · KhiladiPro · Production Model

Shuttle.

Production shuttle tracking model for live badminton broadcasts. 93% accuracy on held-out clips from both broadcast footage and KourtEye captures, on held-out venues — lifted from 80% via flywheel-driven retraining on annotator-corrected data. Two model variants: TrackNetV5 (heavy, production accuracy) and TrackNetV5Lite (lightweight, INT8 dynamic, on-device deployable at 69.5 FPS).

Production Accuracy
93%
Flywheel Lift
80% → 93%
Lite Variant Speed
69.5 FPS
Status
Live · Serving
01 / The Problem

Shuttles are small, fast, and almost invisible.

A shuttlecock in broadcast badminton footage is often under 8 pixels across. During a smash, it travels at effective velocities exceeding 300 km/h at the racket — the camera captures it as a blurred streak across multiple frames. The net occludes it every few seconds. And on broadcast lighting, the white shuttle on a light court background becomes nearly indistinguishable from the surface it's flying over.

Tracking that is the foundation of every downstream badminton analytics product — line-call detection, smash speed estimation, rally segmentation, highlight generation, biomechanics. The shuttle tracker is the substrate. If it's wrong, everything downstream inherits the error.

Off-the-shelf TrackNet variants underperformed on the footage KhiladiPro actually ships against — broadcast cameras with varied lighting, KourtEye in-house captures with different angles, venues with court-color variance, shuttles flying through net-occluded paths. The job: take a known-good architecture and make it materially better on the data we actually see.

02 / The Lift

80% to 93% — same architecture, better data.

The accuracy gain didn't come from a new model. It came from a better training pipeline — the KhiladiPro Annotator's model-in-the-loop schema (pred_/man_/edited) produced data quality that compounded across multiple retraining cycles. The architecture stayed; the data improved.

75% 80% 85% 93% ~80% Baseline ~85% Cycle 1 ~89% Cycle 2 ~91% Cycle 3 93% PRODUCTION — FLYWHEEL CYCLES (annotator-corrected retraining) —
Accuracy across retraining cycles. Each cycle = annotator corrections fed back through the model-in-the-loop schema, biased toward failure-mode frames flagged by the annotation_score metric. Same TrackNet architecture across all five evaluations.
03 / Two-Model Strategy

Heavy for accuracy. Lite for deployment.

Production sports CV doesn't pick between accuracy and latency — it ships both, and chooses which one to consume at inference time depending on the consumer. The same training data produces both variants. Same architecture family, different parameter counts and quantization strategies.

— TrackNetV5 (Heavy)

Production accuracy reference

Full-capacity TrackNetV5 architecture. Higher parameter count, server-side inference. This is the model that delivers the 93% headline accuracy against the held-out broadcast + KourtEye evaluation set. Used for high-stakes analytics paths where latency budget is generous — match analysis, batch processing, ground-truth generation for downstream evaluation.

Accuracy
93%
Deployment
Server-side
Where it ships: the model running on production analytics pipelines where the accuracy budget matters more than the latency budget. Reference benchmark for all downstream evaluations.
— TrackNetV5Lite

On-device deployment target

Halved-channel TrackNetV5 architecture, INT8 dynamic quantization. 0.29 MB on disk — small enough to ship as part of an iOS/Android app bundle. Runs at 69.5 FPS on Apple CPU (14.38 ms/frame), no GPU required. Accuracy trade-off vs heavy is bounded — the Lite variant is the deployment target for live mobile capture, edge inference, and bandwidth-constrained contexts.

Size
0.29 MB
Speed
69.5 FPS
Format
INT8 dynamic
Hardware
Apple CPU
Where it ships: mobile captures via the KourtEye app, on-device inference for real-time feedback, and any deployment context where the model has to live inside a binary, not behind an API.
04 / Approach

What actually changed across the cycles.

The 80% → 93% lift wasn't a single architectural breakthrough. Four levers stacked, each one producing a few percentage points of improvement. The flywheel mechanism was what made the levers reproducible.

More & better
training data
The annotator platform produced labelled data at higher velocity than off-the-shelf tools, and crucially — produced corrected data, not just labelled data. The pred_/man_/edited schema preserved the model's predictions alongside human corrections, so the next training cycle could target exactly the frames where the previous model was wrong. The training set didn't just grow; it grew toward failure modes.
FeatherTrack —
lightweight architecture
The halved-channel TrackNetV5Lite variant. Designed for on-device deployment with INT8 dynamic quantization — but the architecture's lower parameter count also reduced overfitting risk on the early training cycles, when the dataset was smaller. The Lite model converged faster and gave us a deployable production target while the heavy model continued improving.
Loss function
tuning
TrackNet's gaussian-heatmap regression loss was tuned to weight high-confidence detections heavier, reducing the false-positive count that dominated earlier evaluations (recall the Nov 2025 benchmark showed 4,726 FP across just 57 visible-ball frames). The loss adjustment biased the model toward precision in addition to recall.
Targeted
augmentation
Augmentation focused on the failure modes the previous cycle's errors surfaced — motion blur, lighting variance, low-contrast backgrounds, partial occlusion. Random augmentation alone doesn't fix specific failure modes; targeted augmentation does. The annotation_score metric flagged which clips the model struggled on, and those failure modes drove the next cycle's augmentation strategy.
05 / Evaluation

Held-out clips. Held-out venues.

The 93% accuracy number is produced on an evaluation set with two layers of held-out separation: (1) held-out clips from both broadcast badminton footage AND KourtEye in-house captures, and (2) held-out venues — venues the model has not seen any training data from. This is the methodology a production sports CV team should use, not the random-split methodology that inflates numbers across portfolio projects.

The reason for both layers of held-out: shuttle appearance varies meaningfully across broadcast configurations (camera angle, exposure, lighting), and venue-specific factors (court color, lighting rigs, background noise) introduce systematic bias that random clip splits can't isolate. If you train on Wankhede footage and test on Wankhede footage, you're testing the venue, not the model.

Accuracy
93%
On held-out broadcast + KourtEye clips, held-out venues.
Lift
+13pp
From 80% baseline via flywheel-driven retraining.
Lite Speed
14.4ms
Per-frame inference on Apple CPU (69.5 FPS).
Lite Size
0.29MB
INT8 dynamic — ships in mobile app bundle.
06 / Honest Limitations

Where the 7% still lives.

— DOMINANT FAILURE MODE

Low-contrast frames are still hard.

When the shuttle's appearance approaches the court or background luminance — white shuttle on light court surface, light shuttle against pale broadcast lighting — the model's confidence drops sharply. This is the single largest contributor to the remaining 7% error.

Low-contrast isn't a problem the model can solve in isolation. It needs more low-contrast training examples (which the annotator is producing), and likely auxiliary inputs in V2 — temporal smoothing across frames, motion-based detection priors, or contrast normalization at preprocessing time.

This is being prioritized for the next flywheel cycle. The annotation_score metric already flags low-contrast clips for upweighting in the next training set.

02
Fast-motion frames during smashes (>50 px/frame movement) remain harder than steady-state rally frames. Earlier evaluations showed motion-blurred frames at 5–10× the typical pixel error. The current model has improved but not eliminated this.
03
Net occlusion — when the shuttle passes through the net plane, the model has to track across frames where the shuttle is partially or fully invisible. Handled today by temporal context but not perfectly.
04
Lite variant accuracy delta vs heavy — the lightweight model carries an accuracy trade-off (currently around ~90% detection rate vs 93% for the heavy variant). For latency-sensitive on-device paths this is acceptable; for analytics paths it's not the right model.
05
Singles only. Most of the training data and evaluation methodology was built around singles matches. Doubles introduces additional complexity (4 players occluding the shuttle, mixed-doubles court coverage variance) and is being addressed by the annotator's doubles clip pipeline.
06
Apple CPU benchmark specificity — the 69.5 FPS number is measured on Apple CPU. Performance on other deployment hardware (Android Snapdragon, x86 server CPUs, ARM) hasn't been formally benchmarked in the same evaluation pass.
07 / What's Next

The next 2–3 percentage points.

Each flywheel cycle has historically delivered 2–4 percentage points of accuracy lift. The next cycle's targeted improvements:

— A note on detail
This page describes the model's design, training methodology, and measured outcomes. Internal training pipelines, dataset compositions, repository paths, and specific deployment infrastructure are not shown — this is production work for KhiladiPro, and the engineering substance is what's discoverable here without exposing IP.