• Role: ML Engineer (Computer Vision) + Backend Developer (demo API)
• Timeframe: Not specified
• Platform: Model training + REST API demo
• Status: Completed (academic project)
• Team: Team project

American Sign Language alphabet recognition from images. The goal was to localize the hand and classify the sign into one of 26 letters. We evaluated multiple object-detection approaches (SSD, YOLOv5, Faster R-CNN), used transfer learning, and built a simple Django API for inference demos.
Key highlights:
• Compared SSD-ResNet50, YOLOv5, and Faster R-CNN pipelines on the same labeled dataset
• Used Roboflow augmentation to scale the dataset and standardize labels
• Shipped a prediction endpoint returning letter, bounding box, and confidence

• Needed both localization (bounding box) and classification (26 classes) with limited labeled data.
• Training compute was constrained early (no GPU), forcing careful iteration choices.
• Offline validation metrics did not translate to real-world video performance.

We framed the task as object detection and used transfer learning to accelerate progress. Data was augmented via Roboflow (rotations and flips) to increase per-class coverage. We then benchmarked SSD-ResNet50 vs YOLOv5 vs Faster R-CNN, focusing on the gap between validation metrics and live inference. For the demo, we wrapped inference behind a Django REST endpoint that accepts an image and returns the predicted letter, bounding box, and score.
• Prioritized a reproducible pipeline over a single “best” model

• Dataset: Kaggle ASL alphabet images (26 classes) with existing annotations
• Preprocessing: Roboflow augmentation + train/val/test split + COCO/YOLO export formats
• Models: transfer learning on SSD-ResNet50 (TF Object Detection Zoo), YOLOv5, Faster R-CNN (Detectron2)
• Training: GPU-enabled runs for longer training schedules; tracked loss and detection metrics
• Evaluation: compared validation metrics vs real-world tests (video inference)
• Demo: Django API endpoint → model inference → JSON response (class, box, score)

• Managed compute constraints: early SSD training was too slow without GPU, forcing a model strategy change
• Diagnosed metric vs reality mismatch: high validation scores but poor live video accuracy
• Built a consistent labeling pipeline across frameworks (TF OD API, YOLO format, COCO JSON)
• Tuned augmentation to increase data volume while monitoring degradation in generalization
• Investigated failure modes separately for box regression vs class prediction (correct box, wrong class)
• Packaged inference into an API that returns structured outputs suitable for a front-end demo

• Produced an end-to-end ASL alphabet detection pipeline with multiple model baselines
• Demonstrated that offline metrics can be misleading without real inference testing
• Delivered a working inference API for demo usage (image in → prediction out)

• Architecture: Transfer learning, object detection (SSD / YOLOv5 / Faster R-CNN)
• Backend/Infra: Django, REST API, Detectron2, TensorFlow Object Detection Zoo
• Tooling: PyTorch, Roboflow, Kaggle dataset, Google Colab/local GPU training