Consumer Goods Image Object Detection
Budget: $250 – $750 USD
I need a robust computer-vision pipeline that can automatically spot and label consumer-goods products in high-resolution images captured on our manufacturing line. The goal is to distinguish each finished item from background equipment, operators, and any other visual noise so that we can feed the detections into our downstream QA and inventory systems.
You will start with raw JPEGs taken under factory lighting, then design, train, and validate an object-detection model—PyTorch or TensorFlow is fine—capable of achieving consistent, real-time performance on an NVIDIA GPU. If you prefer a different framework, let me know why; I’m flexible as long as the final solution is easy for my engineering team to maintain.
Deliverables
• Annotated sample dataset (in COCO or YOLO format)
• Trained detection model with weights and versioned codebase
• Inference script or REST microservice that returns bounding-box coordinates, confidence scores, and class labels for each detected consumer-goods product
• Brief deployment guide so my team can replicate results on their own workstation or server
Acceptance criteria
1. Mean Average Precision (mAP) ≥ 0.85 on a held-out test set I’ll provide.
2. Inference time ≤ 120 ms per 1920×1080 image on an RTX 3060.
3. Clean, well-commented Python code that installs with a single requirements.txt or environment.yml.
If you have existing models that can be fine-tuned on our data, mention that in your proposal along with any relevant benchmarks you’ve achieved on similar manufacturing scenarios.
You will start with raw JPEGs taken under factory lighting, then design, train, and validate an object-detection model—PyTorch or TensorFlow is fine—capable of achieving consistent, real-time performance on an NVIDIA GPU. If you prefer a different framework, let me know why; I’m flexible as long as the final solution is easy for my engineering team to maintain.
Deliverables
• Annotated sample dataset (in COCO or YOLO format)
• Trained detection model with weights and versioned codebase
• Inference script or REST microservice that returns bounding-box coordinates, confidence scores, and class labels for each detected consumer-goods product
• Brief deployment guide so my team can replicate results on their own workstation or server
Acceptance criteria
1. Mean Average Precision (mAP) ≥ 0.85 on a held-out test set I’ll provide.
2. Inference time ≤ 120 ms per 1920×1080 image on an RTX 3060.
3. Clean, well-commented Python code that installs with a single requirements.txt or environment.yml.
If you have existing models that can be fine-tuned on our data, mention that in your proposal along with any relevant benchmarks you’ve achieved on similar manufacturing scenarios.