TriangleNet: Edge Prior Augmented Network for Semantic Segmentation through Cross-Task Consistency

Cityscapes

• The expriments were conducted using single V100 GPU with batch size 4.

Model	Backbone	Resolution	Training Iters	mIoU(val)	mIoU(test)	Links
Baseline	ResNet-18	1024x1024	300000	76.77	74.48	model
Trianglenet	ResNet-18	1024x1024	300000	78.96	77.36	model

comparision with real-time semantic segmentation models

• All these inference speeds were measured using PaddleInference Api on a A100 GPU device. During this process, we use the PaddlePaddle implementations of these state-of-the-art models provided by PaddleSeg for fair comparision.

Model	Backbone	Test Resolution	mIoU(test)	FPS
ESPNetV2	-	512x1024	66.2	126.5
BiSeNetV1-L	ResNet18	768x1536	74.7	83.9
STDC1-Seg50	STDC1	512x1024	71.9	262.1
STDC2-Seg50	STDC2	512x1024	73.4	207.4
STDC1-Seg75	STDC1	768x1536	75.3	152.7
STDC2-Seg75	STDC2	768x1536	76.8	131.5
PP-LiteSeg-T1	STDC1	512x1024	72.0	219.4
PP-LiteSeg-B1	STDC2	512x1024	73.9	184.3
PP-LiteSeg-T2	STDC1	768x1536	74.9	141.2
PP-LiteSeg-B2	STDC2	768x1536	77.5	118.4
TriangleNet	ResNet18	1024x2048	77.4	46.2

FloodNet

• The expriments were conducted using 4 V100 GPUs with batch size 16.

Model	Backbone	Resolution	Training Iters	mIoU(test)
Baseline	ResNet-18	1024x1024	20000	65.64
TriangleNet	ResNet-18	1024x1024	20000	70.97