junbin666/tiny-vial-cnn-sdg-objectness-v2

Tiny Vial CNN SDG Objectness v2

Tiny CNN v2 adds an explicit objectness head to the v1 bounding-box baseline. It can return “no object” for negative synthetic frames before producing a single blue_vial box. It supports the public Isaac Sim-to-Real VLA Lab and uses junbin666/isaac-blue-vial-sdg-contract.

This is an educational synthetic-data checkpoint, not a production detector or safety claim.

Test results

Metric	Result
Precision	1.000
Recall	1.000
False-positive rate	0.000
Accuracy	1.000
Positive mean IoU	0.849587
IoU ≥ 0.5 on positive frames	1.000

The model-specific test holdout has only 11 frames: 9 positive and 2 negative. Therefore, 0.0 FPR means 0/2 false positives, not evidence of production robustness. The public dataset repository keeps its original 81/10/9 split; for v2 objectness training, the 20 source negatives are deterministically divided 16/2/2 across train/validation/test. Positive frames retain their original source splits.

Architecture

• 64×64 RGB input
• Three stride-2 convolution layers: 3→12→24→32
• Shared 64-unit representation
• One objectness logit plus four sigmoid box values
• BCE-with-logits objectness loss
• Smooth L1 box loss, applied only to positive samples
• 141,357 parameters

Files

• model.safetensors: recommended state dictionary
• pytorch_model.pt: original checkpoint, load with weights_only=True
• config.json: preprocessing, threshold, and output contract
• inference.py: standalone positive/negative inference
• eval_results.json: confusion counts and per-frame held-out results
• training_metrics.jsonl: deterministic training history
• release_report.json: hashes and package verification

Inference

pip install -r requirements.txt
python inference.py examples/positive-000006.png --weights model.safetensors --output prediction.png
python inference.py examples/negative-000090.png --weights model.safetensors --output negative.png

The JSON output includes detected, objectness_score, threshold, and an optional pixel-space bbox.

Limitations

The training and evaluation images come from a small procedural renderer. The checkpoint has not been tested on a real camera, Isaac Replicator output, occlusion, multiple objects, unfamiliar containers, or safety-critical robot motion. A robust release needs a substantially larger held-out negative set, real-camera domain-shift evaluation, and calibrated uncertainty.

License

Model code and weights use the MIT License. The associated dataset uses CC BY 4.0.