TR-HASH-Vision-v6-1M-Objects365
Training in progress. This private draft intentionally contains no final weights or release metrics yet.
TR-Hash Vision v6 is a compact anchor-free detector built on an ID-hash-routed MoE vision tower. Spatial-token IDs select expert parameter subspaces while shared attention preserves contextual feature mixing. The detector combines native hierarchical P3/P4/P5 features, shifted-window attention, a lightweight PAN, an optional P2 small-object path, dynamic one-to-many assignment with STAL, decoupled LTRB/DFL regression, unified QFL quality-class scores, and an optional one-to-one NMS-free inference branch. The realized model has approximately 1.31M parameters and uses 640 px inputs on Objects365 v2.
The specialized detector path adds residual adapters at every pyramid level, an ID-hash-routed class x level gate, class x object-size x scene-density weighting, level-balancing and gate-calibration objectives, and supervised object contrastive learning. Motion features are not applied to static images; they are enabled only for an explicit temporal video input branch.
Evaluation
Training is currently in progress; validated metrics will be added here.
Inference
Inference instructions will be added when validated v6 weights are uploaded.
Training
- Dataset: Objects365 v2
- Optimizer: MuSGD
- Routed-expert LR multiplier: 1.5x
- Initialization: Random initialization; no external classification backbone
- Role: Full-detector detection pretraining before controlled COCO refinement
- Training: Streaming-safe flip, color and blur transforms; random-access Mosaic, MixUp and Copy-Paste disabled
- Framework: Complexity Framework
Limitations
This is a research checkpoint under CC BY-NC 4.0. Validate accuracy, calibration, latency, and failure modes on your own target domain. A training configuration or active run is not evidence of accuracy; release claims require a realized checkpoint and an explicit evaluation protocol.