Hive Vision v1.1.0

Hive Vision v1.1.0

Retrained Limelight 3A SSD-MobileNetV2 detector with a significantly larger
and harder dataset, fixing two classes of false positives found in the field.

What's fixed

• Shadows are no longer detected as game elements. Shadow-hardening data
  (shadowed and darkened balls, deep-shadow field footage) was added to the
  training set; shadowed regions are no longer reported as Pollen/Nectar.
• Outside / off-field objects are no longer detected as balls. Hard
  negatives covering wiring/cables, robot chassis, and field signage were
  added; scenes built to trigger these false positives now return zero
  detections.

Model changes

• Retrained SSD
  (yolo/weights/best_limelight3a_ssd_mobilenetv2_300x300.tflite) now ships
  as the default Limelight 3A detector.
• Training dataset grew to ~10,000 labeled images (9,480 train images
  across synthetic renders, real footage, and hard-negative frames), up from
  ~7,000.
• Training run: 20k steps on the Limelight online trainer, final training loss
  0.019, validation loss 0.040 (down from ~0.20 on the previous local run).
• Class order unchanged: yellow_pollen, red_nectar, blue_nectar
  (yolo/weights/labels.txt).

Credit

The shadow and outside-object false positives addressed in this release were
field-reported by Team 16765 ProBotiX (probotixbladel@gmail.com),
Sid — FTC 13100 47 Beavers (Index 47.org), and Javi Nashat
(Lazer Robotics 23286) from their arena and robot testing of the v1.1 model —
thank you for the field reports. Demo footage of the API robot drives was
recorded by ProBotiX and Sid, and the on-robot SSD clip by Javi Nashat.

Fix in action

SSD v1.1.0 on the Fixed Up field scene — all three game elements detected,
nothing else flagged:

And live in the real world, recorded by Javi Nashat (Lazer Robotics 23286):

Clip: demo/javi_v1_1_ssd.mp4

Direct download: javi_v1_1_ssd.mp4

Tooling

• detect_video_ssd_realtime.py now runs the SSD alone — the YOLO/CV ground
  truth and lab overlay were removed, so the viewer reflects exactly what the
  Limelight 3A reports on robot.

Included

• yolo/weights/best_limelight3a_ssd_mobilenetv2_300x300.tflite (SSD v2,
  uint8-in / float-out, 5.0 MB) + labels.txt
• yolo/weights/best_limelight3a_float32.tflite for PC verification
• yolo/weights/best.onnx / best.pt PC YOLO reference
• Annotated MP4 demos and looping GIF previews