Canonical train
1,287 images · 2,959 ground-truth boxes · 199 negative images.
Model engineering · Experimental study
RTMDet-Tiny: from training to ONNX validation.
A one-class detector study documenting training lineage, ONNX runtime profiles, and formal numerical validation.
0.7510mAP50
0.4090mAP50:95
Original Epoch-20 · 283-image Canonical-Val result.
01 / Dataset and scope
One class, fixed validation split.
Canonical validation
283 images · 618 ground-truth boxes · 40 negative images.
Both saved evaluation records cover the same 283 row identities. Cross-stage training-data byte identity: verification pending. Independent held-out test: pending.
02 / Model lineage
Two separate training stages.
Original scratch Pilot · Epoch 20
20 epochs · 6,440 optimizer updates · FP32 · 640×640 · batch 4.
Canonical-Val: 0.7510 mAP50 / 0.4090 mAP50:95.
Separate fine-tuning · Epoch 15 best
Initialized from the original EMA-aligned model state as a separate fine-tuning phase with a new optimizer and scheduler.
Canonical-Val: 0.748 / 0.417 at selected Epoch 15; final Epoch 20 was 0.740 / 0.417. Fine-tuning results are reported on Canonical-Val.
03 / ONNX and network timing
Structural export and network-only profiles.
The original Epoch-20 model was exported as a static FP32 graph with input [1,3,640,640] and six raw heads: cls_p3/p4/p5 and bbox_p3/p4/p5. ONNX structural checking passed. Preprocessing, decoding, filtering, rescaling, and NMS remain outside the graph; raw-head and final-detection parity results are reported below.
| Profile | Device and runtime | Mean | P50 | P95 |
|---|---|---|---|---|
| CPU B2 · 8 / 1 threads | Ryzen 7 5800H · ORT 1.20.1 CPU EP | 71.190 ms | 68.289 ms | 110.230 ms |
| CUDA B3-R1 · 1 / 1 threads | RTX 3070 Laptop · ORT 1.20.2 CUDA EP | 6.212525 ms | 6.093550 ms | 6.881800 ms |
Each profile used 10 warmups and 96 measured calls over the first 32 Canonical-Val indices. GPU timing includes host/device transfer and output return. CPU and CUDA profiles used different providers and ONNX Runtime versions; compare them as separate device profiles. These session.run() measurements cover model-network execution; full-video timing also depends on image I/O, preprocessing, postprocessing, OCR, video work, and application latency.
04 / Numerical validation
Formal parity validation.
Formal D2: FAILED after complete 283/283-image coverage. Frozen acceptance result: FAILED.
Open the stage-by-stage validation record
| Check | Recorded result | Boundary |
|---|---|---|
| Original C1 input parity | PASS · 3-image sample | Input parity passed on the 3-image sample. |
| Original C2 raw-head parity | FAILED · 16/18 comparisons passed | Three elements exceeded frozen 1e-4 tolerance; cause unknown. |
| Formal C3 decode / pre-NMS | PASS · 283/283 | Stage-local C3 pre-NMS decode result. |
| Formal C4 final detections | FAILED · 281/283 | Two threshold-sensitive native NMS decisions failed. |
| Formal D2 overall | FAILED · 283/283 covered | Final outcome: FAILED. |
| Fine-tuned candidate raw heads | FAILED · 9/18 tensors | 96 elements exceeded rtol=atol=1e-4 on a fixed three-image check. |
05 / Limitations
Validation scope and open checks.
- Evaluation set: Canonical-Val. Independent held-out test: pending.
- CPU/CUDA profiles cover the original Epoch-20 graph; fine-tuned candidate profile: pending.
- Fine-tuned candidate Android runtime, quality, and performance A/B: pending. Candidate-selection records currently differ between the research and Android repositories.
- Current-environment reproduction, end-to-end video latency, production acceptance, and training-data byte identity: pending verification.
- Publication rights for source media, labels, checkpoints, and derived ONNX: pending review.