SpeedNet — ego-speed from onboard video (private)
Optical flow → CNN → GRU speed regressor for onboard/POV footage (go-kart,
ATV, car, machinery). ~0.9 M parameters. Trained 2026-07-24 on the private
dataset x2q/onboard-speed-estimation.
Checkpoints
| File | Training | Use when |
|---|---|---|
speednet_v2.pt |
GPS per-frame loss + lap-time weak supervision (Racehall 1000 m laps) mixed from scratch | Recommended. Indoor kart + ATV + low/mid-speed car |
speednet_best.pt |
GPS per-frame loss only | Highway/high-speed car (indoor scale ~2× too low) |
speednet_v3b.pt |
RECOMMENDED. Context branch (RGB frame -> embedding -> GRU) + lap augmentation, trained on 7-domain corpus (own + L2D + comma2k19 + Skagen DK) | First checkpoint best-in-class everywhere: indoor laps 3.4 km/h MAE, ATV 0.65 m/s, highway 5.6 (beats baseline), urban 1.7, Skagen 2.1, L2D 2.7. Needs ctx frame input (see train_v3.py) |
speednet_v2b.pt |
v2 recipe retrained on expanded corpus (own 95 min + 72 L2D EU episodes + 120 comma2k19 segments, fps-normalized flow) | Best all-round GPS domains: ATV 0.73 m/s, L2D 4.2, comma 2.6, highway 8.7; indoor laps 6.0 km/h MAE (−6 bias — lap weight vs bigger corpus, tune LAMBDA to rebalance) |
Test metrics (leave-recording-out)
| Test set | best | v2 |
|---|---|---|
| ATV forest, different day (RMSE) | 0.77 m/s | 0.96 m/s |
| Car urban (RMSE) | 3.6 m/s | 2.3 m/s |
| Car highway 100–135 km/h (RMSE) | 5.8 m/s | 11.9 m/s |
| Indoor kart, holdout laps (lap-mean MAE vs official times) | ~42 km/h | 1.8 km/h (full-lap window) / 4.6 km/h (streaming, +bias) |
Known limitation: monocular scale ambiguity — without scene context the model trades highway accuracy for indoor accuracy (v2).
RGB context branch: TRIED (2026-07-24, v3 concat-embedding and v4 scalar scale-gain). Both partially restored highway but learned session/scene quirks instead of a generalizing scene→scale mapping, degrading other domains. Root cause is data diversity (4 domains, one indoor venue), not architecture — RESOLVED (2026-07-25): with 7 domains (added L2D EU, comma2k19 US, Skagen DK) the context branch (v3b) learned a generalizing scene-scale mapping — single checkpoint best-in-class on all test domains. Lap-forward augmentation also required (prevents lap memorization).
Usage
import torch
from modeling_speednet import SpeedNet, predict_streaming, compute_flow
model = SpeedNet().cuda()
model.load_state_dict(torch.load('speednet_v2.pt'))
flow = compute_flow('clip_320x180_15fps.mp4') # [N, 72, 128, 2]
speed_mps = predict_streaming(model, flow) # per-frame, m/s
Inference MUST be streaming (hidden state carried across the clip, as in
predict_streaming). Windowed inference with reset state collapses
indoor-domain output to ~half scale. Post-smoothing with a ~1 s rolling
median is recommended.
Input contract: 320×180 @ 15 fps video → Farneback flow → resize 128×72. Other resolutions/framerates change the flow distribution and will mis-scale predictions.
Provenance
- GPS labels: Garmin Dash Cam 66W (car), DJI Osmo Action 5 Pro + GPS remote (ATV/miniloader, djmd protobuf, position-derived 10 Hz).
- Lap labels: official SMS-Timing results (Racehall Aarhus, 1000 m track), aligned to helmet-cam video; holdout laps never trained.
- The project's older fused indoor telemetry was rejected as ground truth.