cuibinge commited on
Commit
8ea64b0
·
verified ·
1 Parent(s): cfe4840

Sync strict polygon dataset importer

Browse files
Files changed (1) hide show
  1. README.md +15 -0
README.md CHANGED
@@ -78,6 +78,19 @@ satellites, sensors, resolutions, and fusion states.
78
 
79
  See [docs/dataset_standard.md](docs/dataset_standard.md).
80
 
 
 
 
 
 
 
 
 
 
 
 
 
 
81
  ## Key Scripts
82
 
83
  - `scripts/build_marine_feature_dataset.py`: scan local or server-side data roots
@@ -86,6 +99,8 @@ See [docs/dataset_standard.md](docs/dataset_standard.md).
86
  normalized `hf://` manifest references without downloading full repositories.
87
  - `scripts/import_quality_gated_hf_dataset.py`: import only HF samples that pass
88
  explicit quality adapters; unverified samples stay in review/reject manifests.
 
 
89
  - `scripts/compose_task_profile.py`: compose Markdown capability cards into a
90
  JSON task profile.
91
  - `scripts/infer_whole_scene.py`: sliding-window whole-scene inference with
 
78
 
79
  See [docs/dataset_standard.md](docs/dataset_standard.md).
80
 
81
+ For the SAMPoly-style polygon head, the dataset gate is stricter: only real
82
+ mask or polygon annotations enter the main polygon training set. Bbox-only
83
+ labels are rejected because rectangles cannot supervise true boundaries,
84
+ vertices, or polygon ordering. The polygon-ready layout is:
85
+
86
+ ```text
87
+ images/{train,val,test}
88
+ masks/{train,val,test}
89
+ manifests/accepted_polygon_samples.jsonl
90
+ manifests/rejected_polygon_samples.jsonl
91
+ dataset_card.json
92
+ ```
93
+
94
  ## Key Scripts
95
 
96
  - `scripts/build_marine_feature_dataset.py`: scan local or server-side data roots
 
99
  normalized `hf://` manifest references without downloading full repositories.
100
  - `scripts/import_quality_gated_hf_dataset.py`: import only HF samples that pass
101
  explicit quality adapters; unverified samples stay in review/reject manifests.
102
+ - `scripts/prepare_polygon_dataset.py`: build the strict mask/polygon dataset
103
+ for SAMPoly-style training and reject bbox-only samples.
104
  - `scripts/compose_task_profile.py`: compose Markdown capability cards into a
105
  JSON task profile.
106
  - `scripts/infer_whole_scene.py`: sliding-window whole-scene inference with