All modalities are stored as dense spatial tensor maps (not RGB renders):
- depth/edge/chm: [1,H,W] float16 continuous values
- segm: [1,H,W] uint8 dense class ID map (0-16)
Add training code examples: one-hot encoding, per-class embedding,
aux-encoder injection. Document why tensors over RGB renders (70% info
loss + OOD for DINOv3).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Mixed-size datasets (e.g. GTA-UAV: drone 512x512 + satellite 256x256)
caused torch.stack error in run_edges_stage. Now groups depth tensors
by spatial resolution and processes each group separately.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Verified on 128 satellite images after dark water fix (std threshold
0.08 → 0.18). Document calibrated thresholds and measured improvements.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
GTA-UAV dark ocean images have std ~0.10-0.15 (blue gradient), not
<0.08 as assumed. Raised dark_water_std_thr to 0.18 to cover all
dark water images while staying safe for non-water backgrounds.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- All datasets now use unified 17 classes (not 11/14/16 separately)
- Document seg post-processing (dark water fix, wetland reclassify)
- Update test count (149), time estimates (17 prompts), palette
- Add seg_classes.py to project structure
- Fix outdated references throughout
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
All datasets now use the same 17-class prompt list for transfer
learning compatibility. World-UAV segmentation needs regeneration.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extract shared UNIFIED_PROMPTS (17 classes, ID 0-16) into
scripts/seg_classes.py for transfer learning compatibility.
Both run_gta_uav.py and run_uav_visloc.py now import from it.
Key change: swimming pool moved from ID 13 → ID 16, so sports field
(ID 13), muddy ground (14), embankment (15) have stable IDs across
both datasets. Missing classes in a dataset = 0 pixels = 0 loss.
Updated: README, CLAUDE.md, segmentation_class_analysis.md, palette.
Deleted old UAV_VisLoc segmentations (need regeneration with 17 classes).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add scripts/run_gta_uav.py for GTA-UAV-LR (48K images, GTA V synthetic)
- 14 segmentation classes: 11 base + bare soil, rooftop, swimming pool
- Fix source filter to recognize "satellite" folder (alongside "DB")
- Document GTA-UAV characteristics in segmentation_class_analysis.md
- Update README and CLAUDE.md with GTA-UAV support
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Default 0.3 was too aggressive — dark water, muddy ground, and
uniform textures fell below confidence threshold. Lower to 0.1
to reduce false background pixels.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Background analysis on crop_0_0/0_1 showed wetland/embankment areas
(dark grey-green, avg RGB 101,103,97) between water bodies not covered
by any existing class. Add:
- "muddy ground and wetland" (ID 14)
- "embankment and levee" (ID 15)
Total: 14 → 16 classes for UAV_VisLoc.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Relative 'in/weights' path fails when script is run from outside the
project root. Use _PROJECT_ROOT / "in" / "weights" instead.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
SegConfig(prompts=None) crashes on len(None) when gin config is not
loaded. Pass the 11-class prompt list directly.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Force save_npy=True when save_safetensors=True (lossless intermediate)
- Auto-enable cleanup_npy to remove intermediates after consolidation
- Save segmentation PNG in palette mode "P" (class IDs recoverable)
- Consolidation now works correctly from PNG-only previous runs
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bundle all per-image modalities (depth, edge, chm, segm) into a single
.safetensors file for fast training DataLoader reads (~0.1ms zero-copy mmap
vs ~5ms for 4x PNG). Adds consolidate stage after main pipeline stages,
save_safetensors/cleanup_npy config flags, resume support, and 10 new tests.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
