| --- |
| title: HyperView RESISC45 Geospatial |
| emoji: 🛰️ |
| colorFrom: green |
| colorTo: blue |
| sdk: docker |
| app_port: 7860 |
| pinned: false |
| models: |
| - hyper3labs/hyper3-clip-v0.5 |
| - openai/clip-vit-base-patch32 |
| datasets: |
| - tanganke/resisc45 |
| tags: |
| - hyperview |
| - geospatial |
| - image-retrieval |
| - remote-sensing |
| --- |
| |
| # HyperView - RESISC45 Geospatial Retrieval Comparison |
|
|
| This demo builds a balanced NWPU-RESISC45 remote-sensing subset and opens |
| HyperView with two pinned scatter panels plus a compact scene-retrieval readout: |
|
|
| - CLIP ViT-B/32 in a Euclidean 2D layout |
| - Hyper3-CLIP `hyper3-clip-v0.5` from `hyper-models` in a Poincare 2D layout |
|
|
| The readout panel ranks live query examples by where Hyper3-CLIP improves the |
| top-10 neighborhood over CLIP. The demo is designed around remote-sensing review |
| workflows: semantic search over aerial/satellite tiles, scene-neighborhood |
| inspection, and label-QA for coarse operational groups such as transport, |
| built environment, agriculture/vegetation, water/coastal, sports/recreation, |
| and natural terrain. |
|
|
| ## Industry Story |
|
|
| Use this as the non-retail proof point for Hyper3-CLIP. The buyer-facing story |
| is not generic satellite classification; it is retrieval QA for large imagery |
| archives: |
|
|
| - Search for one scene tile and inspect whether its nearest neighbors stay in |
| the same exact scene class and broader operational group. |
| - Compare CLIP and Hyper3-CLIP on the same query to surface scene drift, |
| mixed neighborhoods, and coarse-label leakage. |
| - Map the workflow to geospatial search, agriculture monitoring, insurance and |
| risk review, infrastructure analytics, and remote-sensing dataset QA. |
|
|
| The demo should be positioned as a hierarchy-sensitive retrieval probe for |
| remote-sensing imagery, not as a replacement for specialist remote-sensing |
| classifiers. |
|
|
| The runtime sample is drawn from `tanganke/resisc45`, using the `test` split by |
| default and selecting a balanced curated subset of RESISC45 classes that support |
| the landing-page comparison queries. Override `GEOSPATIAL_DEMO_LABELS` to change |
| the comma-separated class list. |
|
|
| Run locally from the HyperView repo: |
|
|
| ```bash |
| uv run python hyperview-spaces/spaces/geospatial-eurosat-clip-hyper3clip/demo.py |
| ``` |
|
|
| Useful overrides: |
|
|
| ```bash |
| HYPERVIEW_PORT=6264 GEOSPATIAL_SAMPLES_PER_CLASS=5 \ |
| uv run python hyperview-spaces/spaces/geospatial-eurosat-clip-hyper3clip/demo.py |
| ``` |
|
|
| ## Benchmark Context |
|
|
| Internal lightweight retrieval runs currently show Hyper3-CLIP above OpenAI |
| CLIP-B/32 on NWPU-RESISC45 same-class retrieval and coarse parent grouping. |
| Keep the public claim narrow: this is a positive geospatial retrieval probe |
| against OpenAI CLIP-B/32, not a claim that Hyper3-CLIP beats every |
| remote-sensing model. |
|
|
| ## Deploy Source |
|
|
| This folder deploys to `mnm-matin/HyperView-EuroSAT-Geospatial` as the live |
| Space. The `hyper3labs/HyperView-EuroSAT-Geospatial` Space is kept as a |
| commit-synced org mirror, but remains paused. |
|
|
| Hyper3-CLIP weights are loaded through the `hyper-models` catalog entry for the |
| gated `hyper3labs/hyper3-clip-v0.5` model repository at runtime. The Space needs |
| an `HF_TOKEN` secret with access to that model. |
|
|
