TII OrbitSight Challenge

Criteria for Application

To participate, participants are required to:

• Register on the challenge platform
• Complete the participation form
• Docker image
  • What to upload: a Docker image exported with docker save (e.g., docker save yourimage:tag -o image.tar, or gzip-compressed as image.tar.gz).
  • How it runs: the container must run non-interactively and finish on its own (no manual input).
  • Read inputs from: /OrbitSight_dataset (mounted read-only) — contains the NVS event recordings (training and testing sequences), ground-truth metadata – in *.txt, instructions on how to check, load, and start with the data using the provided /OrbitSight_dataloader folder. Model files: include your AI model weights, model structure file, and inference script inside the Docker image. The inference script must accept the OrbitSight event recordings (training and testing sequences in *.npy) as input and produce a .txt prediction results file within the container.
  • Write outputs / Result files to: /work/teamName/DDMMYYYY (mounted write-only) — the Portal collects this folder to get the prediction results and scoring sheet. Save all outputs inside this folder — one detection per row using the following fields: sequence_id, timestamp_us, x, y, w, h, class_id, confidence. Files should be saved as <sequencename>.txt, along with an Evaluation_Metrics.xlsx file inside the same folder.
  • Entry command: provide an automatic entrypoint (e.g., CMD ["sh","run.sh"]). Avoid interactive shells.
  • Network: containers run offline (no internet access).
• Submit a 5-page technical proposal (in .pdf) covering the following:
  • Problem statement and proposed solution
  • Outcome metrics: Effectiveness measures used to evaluate the solution (mAP (mean Average Precision), precision, recall, and F1 score, inference efficiency)
  • Value proposition and competitive positioning
  • Technical approach and solution architecture: Detailed methodology and expected outputs
  • Team capacity: Participant background, capacity, and capability to develop the solution
  • Prior work: Details on any proof of concept (POC), additional development, or existing applications of the solution

Submission Deadline: 9 September at 23:59 (GMT+4)

All deliverables must be submitted in English.

Reproducibility verification (top 5 finalists only)

After the submission deadline and final scoring are complete, the top 5 ranked teams will be notified privately by the organizers. Within 7 calendar days of notification, these teams must submit:

• The same Docker image and prediction result files submitted in Phase 1, including results obtained from the new unseen dataset provided.
• The full training source code (zipped repository), including model definition, training script, and configuration files used to produce the submitted weights.
• A README.md documenting the training environment (Python version, key dependencies and versions, GPU/CPU hardware used) and the exact command(s) to reproduce training from scratch.
• The training dataset specification (which datasets were used, any preprocessing or augmentation applied).

Solution evaluation hardware specifications

• Hardware (to test): Intel Core i9-12900H CPU (14 cores / 28 threads), 32 GB RAM, Ubuntu 22.04/24.04 LTS. CPU-only execution (no GPU); used uniformly across all submissions for fairness and reproducibility.
• Network: containers run offline (no internet access).