Dear Zindi Team,

We are reaching out to propose a new data science challenge to be hosted on your platform: the Few-Shot Industrial Screen Defect Detection Challenge.

🔍 Challenge Overview:

This challenge focuses on detecting defects in industrial screens from high-resolution images with very limited labeled samples per defect category (few-shot setting). The goal is to push the limits of object detection algorithms in low-data regimes, relevant to real-world industrial inspection scenarios.

📁 Data Description:

• The dataset consists of high-resolution RGB images of industrial display panels.
• Each image has one or more annotated screen defects (e.g., bright spot, line defect, dead pixel).
• Annotations are provided in Pascal VOC XML format (bounding boxes + class labels).

🎯 Task:

Participants must build models that can detect and classify defects in the test set. Due to the few-shot setup, only a small labeled training set will be provided.

🧪 Evaluation Metric:

• Mean Average Precision (mAP) at IoU 0.95 (mAP@0.95) following COCO-style evaluation, but simplified to a single IoU threshold.

📦 Files Provided:

• train/: Images with annotations (XML)
• test/: Images only (no annotations)
• sample_submission.csv: Submission format with image_id, class, bbox (x_min, y_min, x_max, y_max), and confidence score
• label_map.txt: Mapping from class ID to class name

🏆 Rewards:

We are open to discussing prize allocation with your team. Ideally, we plan to sponsor monetary prizes for the top 3 teams.

🗓 Suggested Timeline:

• Challenge Launch: [e.g. June 1, 2025]
• Duration: 6–8 weeks

Please let us know what additional information or materials you would need from us to proceed. We are excited to collaborate with Zindi on this!