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).
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.
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.
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.
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.
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.
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.
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.
4,726 FP across just 57 visible-ball frames). The loss adjustment biased the model toward precision in addition to recall.
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.
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.
~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.Each flywheel cycle has historically delivered 2–4 percentage points of accuracy lift. The next cycle's targeted improvements: