OpenExploer/unet_mobilenetv1

UNet (MobileNetV1)

UNet encoder-decoder structure: MobileNetV1 extracts multi-scale features (stride 4/8/16/32/64); DwUnet upsamples level-by-level with depthwise separable convolutions and fuses feature maps; SegHead outputs segmentation logits at 5 scales; SoftmaxFocalLoss provides joint supervision; at deployment, SegHead exports only the highest-resolution (stride 4) scale argmax result.

---

Deployment Metrics

Model Parameters

Model	Model Input	Backbone	Neck	Model Output
UNet	Single image 1x3x1024x2048	MobileNetV1	DwUnet	Segmentation mask (B,19,H,W)

Accuracy Metrics

March	Metric	float	calibration	qat	hbm
J6M	MeanIOU	0.6774	0.646	0.6707	0.6721

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	1.13	1115.63	15.80
J6P	0.84	6147.35	15.80
J6B	3.61	327.17	13.00

---

Model Overview

Core Design

UNet encoder-decoder structure: MobileNetV1 extracts multi-scale features (stride 4/8/16/32/64); DwUnet upsamples level-by-level with depthwise separable convolutions and fuses feature maps; SegHead outputs segmentation logits at 5 scales; SoftmaxFocalLoss provides joint supervision; at deployment, SegHead exports only the highest-resolution (stride 4) scale argmax result.

• Task type: Semantic segmentation (Semantic Segmentation).
• backbone: MobileNetV1 (alpha=0.25, include_top=False removes classification head, lightweight depthwise separable convolution backbone).
• neck: DwUnet (base_channels=8, output_scales=(4,8,16,32,64), depthwise separable convolution U-Net decoder).
• Segmentation head: SegHead (num_classes=19, outputs at stride 4/8/16/32/64 (5 scales); training outputs logits, deployment outputs argmax).
• Loss function: SoftmaxFocalLoss (scale-weighted, mitigates class imbalance).
• Normalization: BGR→YUV444 then Normalize(mean=128, std=128), then split into multi-scale pyramid input via scale_factors.
• Model input: 1×3×1024×2048 (RGB, native resolution).
• Model output: Per-pixel 19-class semantic segmentation map (B,19,H,W).

Official Repo and Paper

Official repo: https://github.com/pytorch/vision (MobileNetV1 PyTorch implementation) Paper: https://arxiv.org/abs/1704.04861

Note: UNet paper at https://arxiv.org/abs/1505.04597.