Course Project: Computer Vision (M.Sc. Computer Engineering, Unical) Authors: Anastasia Martucci & Giuseppe Zappia
progetto-cv-martucci-271316-zappia-268784.ipynb: The complete training and inference pipeline, including data augmentation, custom loss implementation, and evaluation logic.PRESENTAZIONE_MARTUCCI_ZAPPIA.pdf: Project presentation slides detailing the theoretical background, architectural choices, and experimental results.
This project addresses the challenging task of detecting and segmenting power lines (cables) in aerial images captured by UAVs, based on the TTPLA dataset.
Power lines are "thin objects" (often 1-3 pixels wide) set against cluttered backgrounds (vegetation, urban areas). Traditional segmentation architectures often fail to preserve the continuity of these lines due to aggressive downsampling.
Our Solution: We implemented a custom Instance Segmentation pipeline based on Detectron2 and PointRend, enhanced with a hybrid loss function and a geometric post-processing strategy to recover pixel-perfect details.
(Example of model output: The model successfully segments thin wires against complex backgrounds)
Standard segmentation networks (like Mask R-CNN) project features onto a fixed low-resolution grid (e.g.,
- The Issue: For a cable that is 1 pixel wide, this projection mixes the cable features with the background, causing the signal to "vanish" or become fragmented.
- Our Approach (PointRend): We utilized Point-based Rendering. Instead of predicting all pixels on a fixed grid, the model iteratively selects the most "uncertain" points (the edges) and re-computes their features using high-resolution maps from the FPN (Feature Pyramid Network).
The model extends the Mask R-CNN framework by replacing the standard mask head with a PointRend module. This allows for high-resolution segmentation without the computational cost of processing a fine grid over the entire image.
(Complete pipeline: ResNet-50+FPN Backbone feeding into the PointRend Head for refined segmentation)
- Backbone: ResNet-50 with Feature Pyramid Network (FPN) for multi-scale feature extraction.
- PointRend Head: Iteratively refines the segmentation boundaries by selecting the most uncertain points from the coarse prediction.
- Training Schedule: 3x Schedule (~45k iterations) to ensure optimal convergence.
Given the extreme class imbalance (cables represent <1% of pixels), a standard loss function would fail. We implemented a custom hybrid loss:
- Focal Loss: Focuses the network on "hard" examples (the thin cables), down-weighting the easy background.
- Dice Loss: Optimizes the global overlap (IoU), ensuring the predicted lines are continuous rather than fragmented.
To handle the limited dataset size (842 training images), we designed a two-stage augmentation pipeline:
- Offline (Static): Geometric transformations (Rotation 90°/180°, Flips) to expand the dataset 5x without introducing interpolation artifacts on the thin lines.
- Online (On-the-fly): Photometric distortions (brightness, contrast) and Multi-scale training applied during each epoch.
The task required outputting not just the mask, but the polar coordinates
- Bias Correction: We identified a systematic geometric shift. We applied a systematic correction of
(+1px, +1px)to all predictions, maximizing the Intersection over Union (IoU). - Line Regression: We utilized
cv2.fitLine(L2 Distance) on the segmented pixels to extract the precise angle and distance of the cables.
The model was evaluated using the Line Detection Score (LDS), which combines segmentation quality (mAP, mAR) with geometric angular precision.
- LDS Score Achieved: 2.54.
- Observations: The high score confirms that the predicted lines are not only visually close but geometrically faithful to the physical cables, with high angular precision.
- Anastasia Martucci - GitHub Profile
- Giuseppe Zappia - GitHub Profile
- Detectron2 Library
- TTPLA Dataset
- Original Presentation and Code included in this repository.