Comparative Analysis

Voice-over: ../assets/audio/analysis.mp3

This page summarizes what prior research reports about when different methods work well, where they fail, and what trade-offs to expect. These are literature-based observations, not new experiments.

Classical CV (Edges/Lines/Periodicity)

  • Strengths: fast, simple, interpretable; runs on limited hardware.
  • Works best: clean paint, good contrast, daylight.
  • Weaknesses: fails under wear, shadows, occlusion, or unusual designs [1].

Deep Learning (Detectors/Segmentation)

  • Strengths: better tolerance to viewpoint changes and partial occlusion.
  • Works best: when trained with diverse data; clear daytime scenes.
  • Weaknesses: domain shift (new cities/night), compute for real-time; needs careful evaluation of FPS/latency [2] [7].

VLM / Multimodal (Context & Reasoning)

  • Strengths: adds scene context and explanations (e.g., is it safe to cross?).
  • Works best: as a complement to fast detectors—reasoning about tricky cases.
  • Weaknesses: latency/compute, reliability of explanations, sensitivity to prompts [3] [5].

Key Takeaways

Quick Quiz (5 Questions)

1. Which factor most degrades classical edge/line detectors?





2. Which metric is standard for segmentation?





3. VLMs primarily add which capability?





4. A key deployment metric on embedded devices:





5. Which remains an open challenge?





References on this page: [1] [2] [3] [5] [7]