OpenExplorer/bev_sparse_henet_tinym

SparseBEV + HENet-tiny

SparseBEV performs 3D detection in BEV space via sparse queries: HENet-tiny extracts multi-view camera features, MMFPN fuses multi-scale features, DenseDepthNet uses lidar point clouds for dense depth supervision during training; SparseBEVHead refines sparse queries layer by layer via MemoryBank + SparseBEVEncoder (with AsymmetricFFN, DeformableFeatureAggregation, SparseBEVRefinementModule) and outputs detection boxes.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
SparseBEV	6-camera multi-view images (B,6,3,256,704) + lidar point cloud (B,N,5)	HENet-tiny	MMFPN + DenseDepthNet	3D bounding boxes (B,N,cls+reg)

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	NDS	0.5414	0.5319	—	0.5307
	mAP	0.4338	0.4265	—	0.4235

Results measured with march = March.NASH_M (J6M) configuration; this task has no QAT stage (qat column is —).

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

Performance Metrics

Performance benchmark: FPS is measured with single-core 8 threads; latency is single-core single-thread; memory is peak DDR usage.

March	latency (ms)	fps	Memory Usage
J6M	10.44	98.74	74.70
J6P	7.39	524.54	87.80
J6B	28.57	36.16	72.00

---

Model Overview

Core Design

SparseBEV performs 3D detection in BEV space via sparse queries: HENet-tiny extracts multi-view camera features, MMFPN fuses multi-scale features, DenseDepthNet uses lidar point clouds for dense depth supervision during training; SparseBEVHead refines sparse queries layer by layer via MemoryBank + SparseBEVEncoder (with AsymmetricFFN, DeformableFeatureAggregation, SparseBEVRefinementModule) and outputs detection boxes.

• Task type: BEV 3D object detection (BEV 3D Object Detection).
• backbone: HENet-tiny (type=HENet, in_channels=3, 4 stages [4,3,8,6] blocks, embed_dims=[64,128,192,384], GroupDWCB/AltDWCB/DWCB blocks, S2DDown downsampling; multi-view camera feature extraction).
• neck: MMFPN (in_strides=[2,4,8,16,32]→out_strides=[4,8,16,32], multi-scale fusion) + DenseDepthNet (dense depth supervision from lidar point clouds during training, num_depth_layers=3).
• Detection head: SparseBEVHead (MemoryBank + SparseBEVEncoder + SparseBEVEncoder, with AsymmetricFFN, DeformableFeatureAggregation, SparseBEVRefinementModule, layer-wise sparse query refinement, num_classes=10, num_decoder=6, num_anchors=384).
• Loss function: FocalLoss (cls) + L1Loss (reg) + CrossEntropyLoss (cns) + GaussianFocalLoss (yns).
• Model input: 6-camera multi-view images (B,6,3,256,704) + lidar point cloud (B,N,5) (load_dim=5, use_dim=[0,1,2,4], num_lidar_sweeps=0, dense depth GT auxiliary).
• Model output: 10-class 3D bounding boxes (num_classes=10), (B,N,cls+reg).

Official Repo and Paper

Official repo: https://github.com/MCG-NJU/SparseBEV Paper: https://arxiv.org/abs/2308.09244

Note: The camera backbone HENet is a HEAL in-house implementation; the official repo uses a different backbone.

Reference

For more J6 chip deployment details, see https://developer.horizon.auto/blog/14074