Pipette Angle Challenge¶
The Problem¶
Predict the relative angle between a pipette and a tube from video frames captured by Transfyr's Observatory.
Input: 5 still images + 2 video clips (2 seconds each, 30 fps, 1920x1080)
Expected Output: CSV with filename and predicted angle for each frame.
Key Finding¶
Not enough data
The provided dataset is insufficient for reliable angle prediction—whether training a custom model or using pretrained feature extractors.
The core issue: features from the same class but different videos have similarity scores nearly as low as features from different classes entirely.
Documentation¶
| Page | Description |
|---|---|
| Approach | How I tackled the problem: labeling, feature extraction, similarity analysis |
| Findings | Why the data is insufficient and what would be needed |
| Alternatives | Better solutions: fiducial markers and IMUs |
| Analysis Notebook | Full code and visualizations (Jupyter notebook) |
Run Locally¶
git clone https://github.com/MishaRubanov/transfyr-pipette-challenge
cd transfyr-pipette-challenge
make install && make run
See the GitHub README for details.
AI/ML Usage¶
- Claude (Anthropic) — Code generation, refactoring, documentation
- ResNet-50 / ConvNeXt-Base — Pretrained feature extraction
- Grad-CAM — Model attention visualization