Assembly Line Action Recognition
Budget: $25 – $50 USD
I am seeking a candidate with extensive experience in Computer Vision. Please review the job description carefully and share a preliminary approach when you reach out.
I have already gathered and labeled video of our production line and now want a computer-vision solution that can recognise each step in the process down to specific hand movements, then output the ordered sequence of those actions in real time (or near real time). The goal is simple: confirm that every operator follows the prescribed workflow and immediately flag deviations for supervisory review.
The dataset includes multi-angle RGB footage captured at 30 fps plus the corresponding time-stamped labels for every micro-action—e.g., “pick screw,” “insert left pin,” “tighten with torque-limited driver.” You are free to decide the most effective architecture (two-stream CNN, 3D-CNN, transformer-based, skeleton/pose-based, etc.). I am comfortable with mainstream frameworks such as PyTorch or TensorFlow, and the final code should run on our existing RTX-series GPU workstation.
Deliverables
• A trained model capable of recognising all labelled actions with high precision, outputting their sequence and timestamps.
• Inference script or REST/GRPC service that accepts a live or recorded video stream and returns a JSON event log.
• Clear documentation covering environment setup, training pipeline, and how to add new actions or retrain.
• Short demo video (screen capture is fine) showing the system processing unseen footage and producing the correct ordered list of actions.
Acceptance criteria
The model must achieve at least 90 % F1 on the held-out portion of my dataset, correctly detect action boundaries within ±0.3 s, and run faster than 0.5× real-time on 1080p video.
I have already gathered and labeled video of our production line and now want a computer-vision solution that can recognise each step in the process down to specific hand movements, then output the ordered sequence of those actions in real time (or near real time). The goal is simple: confirm that every operator follows the prescribed workflow and immediately flag deviations for supervisory review.
The dataset includes multi-angle RGB footage captured at 30 fps plus the corresponding time-stamped labels for every micro-action—e.g., “pick screw,” “insert left pin,” “tighten with torque-limited driver.” You are free to decide the most effective architecture (two-stream CNN, 3D-CNN, transformer-based, skeleton/pose-based, etc.). I am comfortable with mainstream frameworks such as PyTorch or TensorFlow, and the final code should run on our existing RTX-series GPU workstation.
Deliverables
• A trained model capable of recognising all labelled actions with high precision, outputting their sequence and timestamps.
• Inference script or REST/GRPC service that accepts a live or recorded video stream and returns a JSON event log.
• Clear documentation covering environment setup, training pipeline, and how to add new actions or retrain.
• Short demo video (screen capture is fine) showing the system processing unseen footage and producing the correct ordered list of actions.
Acceptance criteria
The model must achieve at least 90 % F1 on the held-out portion of my dataset, correctly detect action boundaries within ±0.3 s, and run faster than 0.5× real-time on 1080p video.
Related categories:
CUDA
Machine Learning (ML)
Neural Networks
Video Processing
Computer Vision
Deep Learning
Object Detection
YOLO