What if AI were the solution to understanding sign language?
summary
A project trains YOLOv7 on an ASL Roboflow dataset of 26 classes (1,728 images) with data augmentation to ~2,300 images and moves training to OVH Cloud GPUs (Tesla V100). Using YOLOv7 with Weights & Biases and OpenCV, precision rises from ~0.80 to ~0.95 after augmentation though validation still shows overlaps and false positives (e.g., face or ceiling labeled as letters). The pipeline exports model weights to object storage and serves a real‑time Streamlit app containerized with Docker, with improved but imperfect performance in open environments.
AI-written summary · about
chapters (17)
- 0:00Introduction
- 1:03AI detection goals
- 3:52ASL dataset
- 5:58YOLO V7 algorithm
- 7:22Real-time detection
- 9:53First model issues
- 11:57Dataset limits
- 13:40Data augmentation
- 16:34Augmented dataset results
- 17:59Cloud training tools
- 21:34Transfer learning with YOLO V7
- 24:04Data processing steps
- 27:17Model comparison with W&B
- 30:52Confusion matrix analysis
- 31:54App deployment
- 35:52Second solution test
- 37:17Summary
related talks
30mBuilding a Sign-to-Speech Prototype with TensorFlow, Pytorch, and DeepStack
21mUsing Computer Vision to Solve Interesting Problems
01h 08mMastering Visual AI with Vision-Language Models & Advanced Evaluation Techniques
01h 14mFoundations of Object Detection from an Expert's Overview
35mHow deep learning can help interpret American Sign Language
04mYOLO in R
18mRecognition and Localization of Parking Signs using Deep Learning
17mVision AI in 2025