README.md

metadata

license: cc-by-4.0
pretty_name: HyperKD-RS
tags:
  - remote-sensing
  - segmentation
  - knowledge-distillation
  - time-series

HyperKD-RS: hypergraphical knowledge-distillation targets for remote sensing

Teacher harvest for distilling a tiny promptable segmenter (PicoSAM3) along exact hypergraph edges, plus the multispectral series the neural codec ("Planet-JEPA") trains on. Built for the OpenCV AI Competition 2026 entry Places Not Pixels (Team Picasso).

Unlike HyperKD (MobileSAM2), which infers hyperedges from embedding distances, the hyperedges here are exact:

• temporal — Hungarian identity tracks across up to 41 annual passes (1985–2025); each track collapsed into a per-pixel majority-vote consensus mask that serves as the training target,
• granularity — all three SAM scale masks (s0 subpart / s1 part / s2 whole) per prompt, QC-gated per scale, plus laminar containment parents between chosen-scale masks.

Source

Landsat 4/5/7/8/9 (Microsoft Planetary Computer STAC), 4.8 km windows at 30 m, best annual pass per spot, SLC-off striping screened out (≤5% nodata). Teacher: SAM 2.1 hiera-base-plus (int8), grid-prompted at 8×8 + half-stride densify — the production tokenizer of the entry.

Layout

One shard per AOI: shards/shard_<tag>.npz

array	shape	meaning
images	(S,256,256,3) u8	display RGB per pass
multispec	(S,160,160,4) f16	raw reflectance r,g,b,nir08
dates, platforms	(S,) str	exact acquisition date, platform
masks	(M,256,256) u8	all QC-passing scale masks + consensus masks
samples	(N,11) f32	one row per accepted prompt

samples columns: scene, cx, cy, s0, s1, s2, chosen, track, parent, target, is_consensus — s0..s2 index into masks (−1 = failed that scale's QC), chosen is the production pick, parent the laminar containment parent within the pass (−1 = top-level), target the distillation target (track consensus where the track has ≥2 observations), cx,cy in the 800×800 prompt frame.

Labels are never baked into geometry: concept annotations (SAM 3.1, calibrated per concept via the taxonomy's visual prompts and thresholds) ship separately as INFERRED facts.

Follow along

The complete pipeline — fetchers, tri-scale harvester, SAM 3.1 labeller, PicoSAM3 distillation harness, taxonomy, AOI manifest, and the QC audit trail — ships in harness/ (also bundled as harness.zip); run order is in its README.

AOIs

32 spots across 12 domains (farmland, grazing, aquaculture, coastal, fire, flood, protected, water, urban/infrastructure, mining, forestry, nuisance) — complete: 32 annual shards (1,217 scenes, 1985–2025) + 6 event shards (_ev, 73 monthly passes inside documented fire/flood windows), every shard with a matching labels/labels_<tag>.npz (SAM 3.1 INFERRED concepts + per-scene cloud masks).