Hive Vision v2.0.0

Hive Vision v2.0.0

v2 of the Limelight 3A ball detector — SSD-MobileNetV2 300×300, fully 8-bit,
max 25 detections per frame. Better than the last model on every axis we
measure: it finds more pollen and nectar in a single frame, it reports fewer
false positives, and it runs with low latency on the camera's CPU.

What's new in v2

• More objects per frame. The model now reports up to 25 detections
  per frame (previous model capped at 10). On a dense ball-pile scene the
  old cap was silently clipping detections — the cap, not the network, was the
  limit. num_detections still tells you how many are real; read slots
  i < num_detections only.
• Fewer false positives. Retrained on a much larger and harder corpus
  (hard negatives: field rig, wiring, signage, deep-shadow frames, off-field
  objects). Lab sweep on the full validation set: 153/347 negative frames
  fired at threshold 0.30, and every active detection scored in [0, 1] with a
  sane box (0 invalid detections).
• Low latency. Full-integer (8-bit) TFLite. Measured on the dev host with
  XNNPACK on the same input: 18–21 ms per inference, versus 113–175 ms
  for the previous float-activation (dynamic-range) build — about 6×
  faster, which is what stops the stream looking choppy on the 3A's CPU-only
  runtime.
• 97.3% precision on the retrain.

Model

File	best_limelight3a_ssd_mobilenetv2_300x300_v2_int8_dets25.tflite
Labels	labels.txt (same folder — upload both)
Architecture	SSD-MobileNetV2, 300 × 300
Input	uint8 [1,300,300,3], scale 1/255
Output	TFLite_Detection_PostProcess — num[1], scores[1,25], class_ids[1,25], boxes[1,25,4] (normalized ymin,xmin,ymax,xmax), float32
Max detections	25
Classes	0 yellow_pollen, 1 red_nectar, 2 blue_nectar
Size	5.19 MB

About max detections = 25: the SSD's non-max-suppression keeps the top-N
highest-scoring boxes after filtering at the score threshold. N=10 meant a
frame with 14 balls could only ever report 10 of them. N=25 gives dense scenes
headroom; sparse frames are unaffected (mean detections per frame is ~4 on the
validation set). Only trust entries where i < num_detections[0] — slots
beyond that are uninitialized memory in the post-process op and can hold
garbage on frames with few detections.

How it was built (custom pipeline)

1. Retrain (V11). SSD-MobileNetV2 300×300 through the TensorFlow Object
   Detection API on the synthetic + real ball corpus, 30,000 steps on an
   RTX 3000 Ada (WSL), class-weighted loss, baked NMS thresholds
   (score 0.3 / IoU 0.6 / per-class 20 / total 20).
2. Export. export_tflite_graph_lib_tf2.export_tflite_model → SavedModel
   → native uint8 input wrapper (u8 × 1/255) → TFLite conversion.
3. Full-int8 quantization. Post-training integer quantization with a
   representative dataset stride-sampled uniformly across the entire train
   set (512 frames, every source prefix represented — taking the first N
   sorted images was an early calibration bug). Weights and activations are
   8-bit; only the post-process outputs are float32, which is the exact
   contract the 3A's neural detector loads.
4. Verification. Contract check (shapes/dtypes/value ranges on active
   slots), latency benchmark, and a full val-set A/B against the previous
   model before it shipped.

Training metrics — V11 retrain

step      30000 / 30000 (100.0%)   TRAIN_DONE
GPU       RTX 3000 Ada (util 95%)

metric	value
classification_loss	0.0395
localization_loss	0.0609
regularization_loss	0.0811
total_loss	0.1815
learning_rate	0.0015
precision	97.3%

Dataset

14,409 labeled images — 11,895 train + 2,514 validation. Synthetic CAD
renders of the game balls, real match/arena footage, and hard-negative frames
(rig, wiring, signage, shadows, off-field objects).

Third-party data used, with license and credit:

Dataset	Author	License
FTC-BioBuzz	Bluedarkup on Roboflow	MIT
Chunc Pollen	Arush's Workspace on Roboflow	CC BY 4.0

Credits

• Team 16765 ProBotiX (probotixbladel@gmail.com) — field testing and demo
  footage.
• Javi Nashat — Lazer Robotics 23286 — field testing and footage.
• Arush — FTC 23511, Seatale Slover — testing, and the Chunc Pollen
  dataset.
• Thank you all — the false-positive reports from the field are what drove the
  negative-data work in this release.

In action

ProBotiX v2 testing (recorded by Team 16765 ProBotiX):

Arush v2 testing (recorded by Arush — FTC 23511):

Install on the Limelight 3A

1. Pipeline type Neural Detector; upload the model and labels.txt
   from this release.

2. Runtime engine: CPU (the 3A has no accelerator — Coral/Hailo settings
   give empty results).

3. Input resolution: 300 × 300.

4. Tuning matters — start from these camera settings and adjust for your
   venue's lighting:

   Setting	Starting value
   Exposure (0.01 ms)	3300
   Black Level Offset	0
   Sensor Gain	25.5
   Red Balance	1266
   Blue Balance	1265

5. Confidence threshold: run at ~80–90% for match play (clean
   detections, no clutter on the stream). Drop it when you want recall over
   precision in practice.

Included in this release

• best_limelight3a_ssd_mobilenetv2_300x300_v2_int8_dets25.tflite — v2 model,
  full-int8, max 25 detections
• labels.txt — yellow_pollen, red_nectar, blue_nectar
• arush_v2_testing.gif, probotix_v2_testing.gif — demo loops

Upgrade notes

• v2 replaces the previous Limelight 3A detector. Class order is unchanged.
• FTC-side decoding is unchanged: iterate getDetectorResults(), or loop
  i < num_detections[0] over the four arrays, skip anything below your
  confidence threshold.