[2026-09] Hyperspectral Object Detection Challenge 2026 — research archive

An OpenKaggle source-and-provenance archive for the Kaggle competition:

• Competition: https://www.kaggle.com/competitions/hyperspectral-object-detection-challenge-2026
• Final deadline: 2026-09-24 16:00 UTC / 2026-09-25 00:00 Asia/Shanghai
• Phase 2 ranking-set release: 2026-09-23 (time not specified on the official page)
• Metric: COCO mAP@[0.5:0.95], with mAP@0.5 reported secondarily
• Daily submission limit: 3
• Maximum team size: 5

What is published

This repository contains the user-authored Kaggle notebook source, smoke tests, validation scripts, experiment metadata, and compact official receipts. It is deliberately not a mirror of the competition workspace.

The following material is excluded:

• all organizer-provided files and competition downloads;
• raw or processed imagery, labels, and submissions;
• model checkpoints, generated predictions, caches, virtual environments, and bytecode.

See DATA_SOURCES.md for the official acquisition path and RELEASE_MANIFEST.md for the publication boundary.

Competition compliance boundary

Competition data remains on the official Kaggle channel. It may only be used for this competition and must not be redistributed, commercially used, cross-channel polluted, or used in a paper without explicit provider permission. Manual test/ranking labels, human prediction of held-out records, private sharing outside the Kaggle team, and any evaluation-system attack are prohibited.

The final prediction must come from one trained detection model. Multiple checkpoints/models may not be combined by voting, weighted fusion, WBF, or post-NMS fusion. The host has explicitly allowed public ImageNet/COCO pretrained weights when model/source/license are declared, and has allowed TTA or multi-scale inference from the same single checkpoint.

Layout

• official/: official receipts, metadata, file inventory, and leaderboard snapshots
• kaggle_notebook/: private Kaggle GPU baseline source and metadata
• scripts/: schema, data, and submission checks
• kaggle_smoke*/: data-mount smoke tests and kernel metadata

First baseline

The initial baseline is a single COCO-pretrained YOLO11m checkpoint trained on deterministic per-image pseudo-RGB made from bands [5, 8, 13], a documented alternative in the official demo. It uses a fixed 80/20 train/validation split with seed 20260909, reports validation mAP, and generates test predictions with one checkpoint only. It is deliberately conservative: no pseudo-labeling, no external remote-sensing teacher, and no model ensemble.