semantic_seg

Cityscapes Semantic Segmentation Evaluation

A comprehensive framework for evaluating semantic segmentation models on the Cityscapes dataset with IoU-based miscoverage categorization.

Project Overview

This project evaluates the SegFormer-B0 model on the Cityscapes validation set for car segmentation (class ID: 13), achieving an mIoU of 0.8688.

Results Summary

Model Performance (SegFormer-B0 on Car Class)

Metric	Value
mIoU	0.8688
Precision	0.9191
Recall	0.9253
F1-Score	0.9155
Images	479

Miscoverage Distribution

Category	Count	Percentage
Excellent (IoU ≥ 0.9)	313	65.3%
Good (IoU ≥ 0.75)	113	23.6%
Over-segmentation	14	2.9%
Under-segmentation	10	2.1%
Critical Failure	7	1.5%
Severe Under-segmentation	6	1.3%
Complete Failure	5	1.0%
Poor Alignment	4	0.8%
Severe Over-segmentation	4	0.8%
Moderate Miscoverage	3	0.6%

Miscoverage Categories：

1. Excellent (IoU ≥ 0.9)

2. Good (IoU ≥ 0.75)

3. Under-segmentation

Model misses parts of the car, resulting in false negatives.

• Causes: Occlusion, poor lighting, small objects

4. Over-segmentation

Model incorrectly predicts non-car pixels as car, resulting in false positives.

• Causes: Similar textures, reflections, shadows

5. Critical/Complete Failure

Severe misalignment or complete failure to detect the car.

• Causes: Extreme weather, heavy occlusion, dataset annotation errors

Error Analysis Color Legend:

• 🟡 Yellow (TP): Correctly predicted car pixels
• 🔴 Red (FP): False positive - incorrectly predicted as car
• 🟢 Green (FN): False negative - missed car pixels