DogFLW: Dog Facial Landmarks in the Wild
This project uses the Dog Facial Landmarks in the Wild (DogFLW) dataset, which provides thousands of dog-face images captured in diverse real-world conditions, annotated with facial landmarks and a face bounding box. The dataset supports tasks such as facial landmark detection, alignment, and downstream facial analysis for dogs.
Mentor Details:
Orna Bregman Amitai
Mentor Details:
Requirments:
- Understand the DogFLW dataset structure and annotations (46 facial landmarks + face bounding box).
- Build and train a model for dog facial landmark detection (regression or heatmap-based).
- Evaluate landmark accuracy and robustness across pose, lighting, and breed variability.
- (Optional) Use detected landmarks for downstream tasks such as alignment, expression/action-unit analysis, or quality control.
Problem Statement
Automated analysis of animal facial expressions and facial structure is limited by the shortage of high-quality, in-the-wild annotated datasets. Dog faces vary widely across breeds, poses, fur patterns, and lighting, making robust facial landmark detection challenging. A strong landmark detector enables applications such as canine facial expression analysis (e.g., DogFACS-related studies), health/welfare monitoring, and improved animal-computer interaction.
Project Objectives
- Understand the DogFLW dataset structure and annotations (46 facial landmarks + face bounding box).
- Build and train a model for dog facial landmark detection (regression or heatmap-based).
- Evaluate landmark accuracy and robustness across pose, lighting, and breed variability.
- (Optional) Use detected landmarks for downstream tasks such as alignment, expression/action-unit analysis, or quality control.
Technical Scope
- Dataset ingestion and parsing of landmark and bounding-box annotations
- Face cropping/alignment and data augmentation (scale/rotation/occlusion)
- Landmark detection modeling (e.g., CNN backbones + heatmap heads)
- Training/validation loops, experiment tracking, and reproducible splits
- Evaluation using landmark error metrics (e.g., normalized mean error) and qualitative visualization overlays
Required Knowledge and Prerequisites
Core Requirements:
- Python programming
- Computer vision basics (image augmentation, coordinate transforms)
- Deep learning framework experience (PyTorch or TensorFlow)
Recommended Background:
- Human/animal landmark detection methods (heatmaps, hourglass/U-Net-style heads)
- Experience with keypoint evaluation metrics and visualization
- Familiarity with domain shift / robustness techniques
Project Difficulty and Expected Level
Overall Difficulty: Beginners
This project is well-suited for:
Teams of 1–3 students
Expected Outcomes
- A reproducible DogFLW training pipeline (data loaders + augmentations + evaluation)
- A baseline dog facial landmark detector with documented performance
- Visualization notebook/app that overlays predicted landmarks and bounding boxes on images
- (Optional) A short report connecting landmark quality to downstream canine facial analysis use-cases