OpenExploer/motr_efficientnetb3

MOTR (EfficientNet-b3)

MOTR models tracking as per-frame query propagation: detection queries are passed across frames via QueryInteractionModule; new targets get new queries and disappeared targets are cleared; training jointly supervises detection and tracking losses.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
MOTR	multi-frame image sequence 1x3x800x1422	EfficientNet-b3	—	tracking boxes + track IDs (B,Q,cls+reg+id)

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	MOTA	0.5837	0.5704	0.5799	0.5767

Data tested with march = March.NASH_M (J6M).

HEAL versions: heal 0.0.2 / hbdk4-compiler 4.11.11 / horizon_plugin_pytorch 3.3.10.

Performance Metrics

Performance test methodology: FPS is measured with 8 threads on a single core; Latency is measured with single core, single thread; Memory is peak DDR usage.

March	Metric	latency (ms)	fps	Memory Usage
J6M	main graph	8.03	128.68	65.30
	qim	0.37	5120.87	6.80
J6P	main graph	5.81	695.38	71.90
	qim	0.36	10348.73	7.00
J6B	main graph	-	-	-
	qim	-	-	-

J6B performance is not available for this model.

---

Model Overview

Core Design

MOTR models tracking as per-frame query propagation: detection queries are passed across frames via QueryInteractionModule; new targets get new queries and disappeared targets are cleared; training jointly supervises detection and tracking losses.

• Task type: Multi-Object Tracking (MOT).
• backbone: EfficientNet-b3 (efficientnet, model_type=b3, include_top=False removes classification head, use_se_block=False, ReLU activation).
• neck: — (no separate neck; backbone features feed directly into MotrHead).
• detection head: MotrHead + MotrDeformableTransformer (d_model=256, num_queries=256, dim_feedforward=1024, Deformable Attention, in_channels=[384]).
• tracking module: QueryInteractionModule (cross-frame query interaction for end-to-end tracking association).
• post-processing: MotrPostProcess.
• Model input: multi-frame image sequence, single-frame size 800 × 1422, sequence sampling interval 10 frames; main graph img:1x3x800x1422, QIM subgraph input is query/feat (no images).
• Model output: per-frame detection boxes + cross-frame track IDs (256 queries, 1 class).

Deployment note: MOTR inference is split into a main graph (backbone + head + post_process) and a QIM subgraph (QueryInteractionModule, DDR input, no images), exported and compiled separately via deploy.py and deploy_qim.py; they work together at inference time (see Usage Guide > Export and Compilation).

Official Repo and Paper

Official repo: https://github.com/megvii-research/MOTR Paper: https://arxiv.org/pdf/2105.03247

Note: backbone is EfficientNet-b3; the official repo uses a different backbone.

Reference

For more J6 chip deployment details, see https://developer.horizon.auto/blog/10353