Real-Time Semantic Segmentation

Problem

Urban scene understanding requires per-pixel classification at interactive speed. Existing state-of-the-art models trade accuracy for latency in ways that make real-world deployment difficult on mid-range hardware.

Approach

We designed a lightweight encoder-decoder with depthwise-separable convolutions and a multi-scale feature pyramid. The decoder uses bilinear upsampling with skip connections to recover spatial detail without heavy computation.

Results

Metric	Value
mIoU (Cityscapes val)	72.4%
FPS (RTX 3080)	38
Parameters	4.2M

Qualitative results show clean boundaries on pedestrians and lane markings even under adverse lighting.

This is a placeholder project entry. Replace with real content in Phase 8.