feature-postprocessing
Guides RoboSat prediction post-processing from probability PNGs to masks, parking GeoJSON features, merge/dedupe cleanup, visual comparison, and vector artifact validation.
- License 许可
- MIT
- 最近更新
- 2026-10-08
- Stars
- ★ 11
- Forks
- 3
01 前言 Front Matter
- name
- feature-postprocessing
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Guides RoboSat prediction post-processing from probability PNGs to masks, parking GeoJSON features, merge/dedupe cleanup, visual comparison, and vector artifact validation.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-reference.md6.8 KB
- 参考文档references/data-formats.md5.4 KB
- 参考文档references/troubleshooting.md7.3 KB
- 参考文档references/workflows.md6.9 KB
03 SKILL.md 正文
RoboSat feature post-processing
Use this sub-skill when a task already has RoboSat prediction probability tiles or mask tiles and needs to turn them into inspectable masks, parking-lot vector features, cleaned GeoJSON, or visual QA mosaics.
Use this for
- Converting binary foreground-probability PNG Slippy Map outputs into palette mask tiles with
rs masks. - Extracting parking-lot WGS84 GeoJSON polygons from mask tiles with
rs features --type parking. - Merging adjacent predicted polygons with a meter threshold via
rs merge. - Deduplicating predictions against existing OSM-derived GeoJSON with an IoU threshold via
rs dedupe. - Building side-by-side image / label / mask mosaics with
rs compare. - Validating post-processing artifacts using bundled scripts.
Route elsewhere
- Need to train a model, load checkpoints, export ONNX, run batch prediction, or produce probability PNGs first: use the
model-lifecyclesub-skill. - Need to extract OSM features before training, cover tiles, download imagery, rasterize labels, compute class weights, or subset tile sets: use the
data-preparationsub-skill. - Need non-parking post-processing handlers: this sub-skill explains the handler contract, but the installed CLI only registers the parking handler by default.
Fast workflow map
- Confirm that probability or mask directories are Slippy Map trees such as
18/69108/105091.png; see data formats. - Convert probabilities to masks:
rs masks out/masks probs/run-a [probs/run-b ...] [--weights 0.7 0.3]. - Convert parking masks to GeoJSON:
rs features out/masks --type parking --dataset dataset.toml out/features.geojson. - Clean vector output:
rs merge out/features.geojson --threshold 3 out/features-merged.geojson, then optionallyrs dedupe osm.geojson out/features-merged.geojson --threshold 0.5 out/features-new.geojson. - Compare visual quality:
rs compare out/compare images labels out/masks [other/masks ...] --minimum 0.01 --maximum 0.95. - Validate artifacts with
scripts/validate_feature_collection.pyand smoke-test soft voting withscripts/softvote_smoke.py.
Detailed command recipes are in workflows. API behavior is summarized in API reference. Common failure modes are in troubleshooting.
Key constraints to remember
- RoboSat probability PNG post-processing is binary: one quantized foreground-probability channel is reconstructed as foreground plus
1 - foregroundbackground. rs featuresis parking-only in the default installed CLI; the dataset config must include theparkingclass and mask class index must match it.ParkingHandlerthresholds are tuned for zoom 18 and will raise on other zoom levels.rs merge --thresholdis measured in meters;rs dedupe --thresholdis intersection-over-union, not distance or model confidence.- Post-processing is CPU-oriented. CUDA is not needed for the commands in this sub-skill.
04 同仓库更多 Skills
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
obsidian-clipper-template-creator
Guide for creating templates for the Obsidian Web Clipper. Use when you want to create a new clipping template, understand available variables, or format clipped content.
python-api
Programmatic Object-Detection-Metrics API guidance for BoundingBox, BoundingBoxes, Evaluator, enums, coordinate conversion, AP/mAP semantics, and source-style import diagnostics.
file-evaluation
Folder-based PASCAL VOC AP/mAP evaluation from ground-truth and detection text files.
object-detection-metrics
Guides legacy Object-Detection-Metrics PASCAL VOC AP/mAP evaluation from detection text folders and direct Python API classes.
nx-workspace-patterns
Configure and optimize Nx monorepo workspaces. Use when setting up Nx, configuring project boundaries, optimizing build caching, or implementing affected commands.