2 of 12

Motivation

TRAIL Lab robot to distribute treats to hikers
Semantic segmentation of hiking trails
Trained on PSPNet with custom dataset (~1000 images)

Baseline Solution

Dataset is overfit to 1 long trail
PSPNet semantic segmentation model is no longer state-of-the-art
Inconsistent results (mIoU)

Limitations

3 of 12

Objectives

Adaptability to perform effectively in various environmental/trail conditions

Robustness

Ability to segment incoming video stream frames in real-time

Speed

Ability to correctly identify and label the trail boundaries

Accuracy

4 of 12

Methodology

Augment data using mirroring, scaling, cropping, and gaussian blurring techniques

Tune the highest performing architecture hyperparameters (epochs, learning rate, batch size)

Data Augmentation

Hyperparameter Tuning

Collect and segment 400 images of various trails �(asphalt, concrete, dirt, gravel)

Explore state-of-the-art segmentation models for optimal performance

Dataset Expansion

Architecture Adjustment

5 of 12

Dataset Expansion

PSPNet + ResNet-50

Percentage of pixels that are accurately classified in the image

Percentage of overlap between the predicted segmentation and the ground truth segmentation

Pixel Accuracy

mIoU

6 of 12

Old Prediction

New Prediction

Trail

Ground Truth

7 of 12

Mirror

Crop

Scale

Gaussian Blur

Data Augmentation

Flip the image along its vertical axis

Extract a portion of the image

Resize the image while maintaining its aspect ratio

Blur the image to reduce noise and suppress detail

8 of 12

Architecture Selection

Performance: How does the model perform on segmentation tasks with existing popular datasets (e.g. Cityscapes or COCO)?

Runtime: Is this approach appropriate for real-time online applications on mobile robot?

Computation Required: How many weights need to be optimized? Is this feasible with only 1 GPU?

PyTorch Compatibility: How complex is it to integrate these models with PyTorch? Are there licensing issues?

9 of 12

Model Results

A Lightweight Encoder-Decoder Network for Real-Time Semantic Segmentation

LEDNet + ResNet-50

PSPNet + ResNet-50

DeepLabV3 + ResNet-50

BiSeNet + ResNet-18

87.79%

86.62%

85.30%

84.09%

82.23%

ICNet + ResNet-50

Performance

10 of 12