OpenExplorer/qcnet_oe

QCNetOE (Trajectory Prediction)

QCNetOE encodes each agent's relative relationships with surrounding map and other agents in a query-centric manner, streaming encoder hidden states agent-by-agent, then the decoder outputs multimodal candidate trajectories and probabilities; removes torch_geometric/torch_cluster dependencies and eliminates most index/gather/scatter ops for quantization-friendly deployment.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
QCNetOE	Streaming scene representation tensor group (B,A,pl,pt,HT)	QCNetOEMapEncoder + QCNetOEAgentEncoderStream	—	Candidate future trajectories (B,A,6,12,2) + trajectory probabilities

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	HitRate	0.8003	0.6817	0.7984	0.7981

Data measured with march = March.NASH_M (J6M) configuration.

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

Performance Metrics

Performance test methodology: FPS is measured with single-core eight-thread; Latency is measured with single-core single-thread; Memory is peak DDR usage.

March	latency (ms)	fps	Memory Usage
J6M	3.72	293.43	34.80
J6P	2.65	1572.60	38.60
J6B	12.25	149.40	32.00

---

Model Overview

Core Design

QCNetOE encodes each agent's relative relationships with surrounding map and other agents in a query-centric manner, streaming encoder hidden states agent-by-agent, then the decoder outputs multimodal candidate trajectories and probabilities; removes torch_geometric/torch_cluster dependencies and eliminates most index/gather/scatter ops for quantization-friendly deployment.

• Task type: Trajectory prediction (Motion Forecasting, multimodal trajectory prediction).
• backbone: QCNetOEMapEncoder + QCNetOEAgentEncoderStream (streaming inference, stream_infer=True; hidden_dim=128, num_heads=8, head_dim=16, num_freq_bands=32, num_map_layers=1, num_agent_layers=1, time_span=2, dropout=0.1; agent-map interaction num_pl2a=32, agent-agent interaction num_a2a=36).
• neck: — (QCNetOE has no standalone neck; encoder output feeds directly into decoder).
• Decoder: QCNetOEDecoder (num_dec_layers=1), outputs num_modes=6 candidate trajectories and their probabilities.
• Preprocessing: QCNetOEPreprocess (stream=True, constructs agent/map relative representations and spatiotemporal relative position encoding).
• Post-processing: QCNetOEPostprocess (output dimension output_dim=2, i.e. predicted trajectory (x, y)).
• Loss: QCNetOELoss.
• Model input: Streaming scene representation tensor group (B=1, A=30 agents, pl=80 map polygons, pt=50 polygon points, HT=10 history steps), including agent/map_polygon/map_point/decoder etc. as OrderedDict inputs; 5s history (num_historical_steps=10) + 6s prediction (num_future_steps=12).
• Model output: 6 candidate future trajectories per agent to predict (num_modes=6, 12 steps, 2-dim coordinates output_dim=2) + probability per trajectory.

Deployment notes: Model supports cold-start/hot-start streaming inference (quant_infer_cold_start controls). Both cali_model and deploy_model enable stream_infer=True; save_memory configured separately for training/deployment (training save_memory=True saves GPU memory, deployment save_memory=False). HBIR export enables enable_vpu=True; compilation uses input_source=["ddr"] (trajectory prediction input read from DDR, not pyramid image input), unlike image-based tasks.

Official Repo and Paper

Official repo: https://github.com/ZikangZhou/QCNet Paper: https://openaccess.thecvf.com/content/CVPR2023/papers/Zhou_Query-Centric_Trajectory_Prediction_CVPR_2023_paper.pdf

Reference

For more J6 chip deployment details, see https://developer.horizon.auto/blog/14089