OpenExplorer/ganet_mixvargenet

GaNet (MixVarGENet)

GaNet models lanes as "root keypoints + along-line offsets": the network predicts an 8× downsampled heatmap to locate lane starting points and regresses per-point offsets to reconstruct the full lane; the decoder filters keypoints by kpt_thr threshold, clusters and merges points on the same lane via cluster_thr, and finally maps back to original image coordinates.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
GaNet	Single frame 1x3x320x800	MixVarGENet	GaNetNeck	Lane point sequences (B,L,P,2)

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	F1Score (F1)	0.7937	0.791	—	0.7908

Results are based on 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 measurement: 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	0.68	2258.92	6.60
J6P	0.59	8623.54	6.80
J6B	-	-	-

J6B performance is not available for this model.

---

Model Overview

Core Design

GaNet models lanes as "root keypoints + along-line offsets": the network predicts an 8× downsampled heatmap to locate lane starting points and regresses per-point offsets to reconstruct the full lane; the decoder filters keypoints by kpt_thr threshold, clusters and merges points on the same lane via cluster_thr, and finally maps back to original image coordinates.

• Task type: Lane detection (Lane Detection).
• backbone: MixVarGENet (include_top=False, output_list=[2, 3, 4] three-scale features, corresponding to stride 8/16/32).
• neck: GaNetNeck: FPN (three-scale fusion to hid_dim=32) + Attention module (attn_ratio=4, position encoding pos_shape=(1,10,25)).
• Detection head: GaNetHead (outputs hid_dim=32 dimensional features).
• Target generation: GaNetTarget (hm_down_scale=8, keypoint radius radius=2).
• Post-processing: GaNetDecoder (root_thr=1, kpt_thr=0.4, cluster_thr=5, downscale=8).
• Loss: GaNetLoss: LaneFastFocalLoss (keypoint classification, weight=1.0) + L1Loss (per-point offset regression weight=0.5 + integer offset regression weight=1.0).
• Model input: Single frame, size 320 × 800 (FixedCrop takes original region (0,270,1640,320) then Resize to 320×800; original resolution 1640×590).
• Model output: Ordered point sequences per lane (original image coordinate system), obtained via GaNetDecoder from heatmap keypoint clustering + offset decoding.

Deployment note: The deployment graph (deploy_model) removes targets/post_process/losses, keeping only backbone + neck + head; output name is pred heatmap features; GaNetDecoder is re-attached during inference (float/hbir/hbm infer) to complete lane decoding.

Official Repo and Paper

Official repo: https://github.com/Wolfwjs/GANet Paper: https://arxiv.org/abs/2204.07335

Note: backbone MixVarGENet is HEAL-developed.

Reference

For more J6 chip deployment details, see https://developer.horizon.auto/blog/14098