Dr Georgios Voulgaris

Dr Georgios Voulgaris

Postdoctoral Researcher in Computer Vision, Deep Learning, and Remote Sensing at the University of Oxford.

Greeks in AI 2025 Symposium — Spotlight Presentation: PerceptiveNet for Tree Crown Semantic Segmentation

Date:

I delivered a single-author spotlight presentation at the Greeks in AI 2025 Symposium, presenting my CVPR EarthVision paper titled “Bridging Classical and Modern Computer Vision: PerceptiveNet for Tree Crown Semantic Segmentation”. The talk introduced PerceptiveNet, a novel backbone architecture designed to address the unique spatial and spectral challenges of dense forest aerial imagery.

Paper:
“Bridging Classical and Modern Computer Vision: PerceptiveNet for Tree Crown Semantic Segmentation”

The presentation covered:

  • Challenges of tree crown semantic segmentation in dense forests, including shadows, occlusions, scale variation, and subtle spectral differences.
  • Limitations of standard CNNs and fixed Gabor filters, including texture bias and restricted adaptability.
  • Introduction of a trainable Log-Gabor-parameterised convolutional layer for improved shape-based and frequency-aware feature extraction.
  • A new backbone architecture combining:
    • a signal-processing-parameterised convolutional layer,
    • Mix Pooling (average and max pooling) for richer spatial statistics,
    • averaged dilated convolutional layers for a wider receptive field.
  • Ablation studies demonstrating the complementary benefits of each architectural component.
  • Integration of PerceptiveNet into a hybrid CNN-Transformer model (PerceptiveNeTr) for long-range dependency modelling.
  • Extensive evaluation across three aerial datasets: TreeCrown, Landcover.AI, and UAVid.

Key findings:

  • Log-Gabor-parameterised convolutional layers outperform standard and Gabor-based layers across all datasets.
  • PerceptiveNet achieves state-of-the-art performance, improving mIoU by 10.5 percent on TreeCrown compared to ResUNet.
  • The architecture generalises strongly to diverse aerial scene segmentation tasks.
  • Class Activation Maps show more focused and discriminative feature extraction compared to standard CNNs.
  • The hybrid PerceptiveNeTr model further enhances performance by capturing global context and long-range dependencies.