MrWangChi/A2PM-SGAD-data

A2PM-SGAD data assets

Precomputed bounding boxes and bbox-count metadata for A2PM evaluation and SGAD / SGAD++ training and region matching evaluation.

• Code and data preparation
• SGAD / SGAD++ checkpoints

Packages

File	Contents	Size
a2pm_evaluation_bboxes.tar.gz	ScanNet1500 and MegaDepth1500 evaluation bounding boxes	1.94 MB
sgad_scannet_bboxes.tar.gz	ScanNet SGAD bounding boxes: 2,689,316 files	389.03 MB
sgad_megadepth_bboxes.tar.gz	MegaDepth SGAD bounding boxes: 394,789 unique files	201.89 MB
sgad_data_metadata.tar.gz	1,887 pair-count sidecars and per-image bbox-count indices	931.16 MB

Sizes use decimal MB. Original images, depth maps, camera data, and LoFTR indices are obtained from the sources linked in the code repository.

A2PM evaluation bounding boxes

a2pm_evaluation_bboxes.tar.gz contains:

• assets/a2pm_bboxes/scannet1500/640x480/: 2,596 files.
• assets/a2pm_bboxes/megadepth1500/1296x968/: 806 files.

Each file contains an N×4 float32 NumPy array with columns x, y, width, height. The directory resolutions describe the bbox coordinate spaces. Different point-matching resolutions share these files.

SGAD / SGAD++ bounding boxes

• ScanNet: sgad/data/scannet_bboxes/train/ (2,477,378 files) and test/ (211,938 files).
• MegaDepth: sgad/data/megadepth_bboxes/train/, with the relative symlink test -> train. Training and test pairs are selected by their respective LoFTR indices.

The bbox arrays use float32 xywh. The data loader applies the configured region filtering. SGAD boxes use their original image coordinate spaces and directory layouts; A2PM benchmark boxes use the evaluation coordinate spaces above.

Bbox-count metadata

sgad_data_metadata.tar.gz contains:

• sgad/assets/pair_bbox_counts/{scannet,megadepth}/{train,test1500}/
• sgad/data/scannet_bboxes/scannet_bboxes_counts.npz
• sgad/data/megadepth_bboxes/megadepth_bboxes_counts.npz

Each sidecar is an [N,2] array in the order of the original scene pair index. Each row gives the bbox counts for the pair's two images. The per-image indices support rebuilding the sidecars.

Download and install

From the code repository root, download and extract the packages needed for your task.

For A2PM benchmark evaluation:

hf download MrWangChi/A2PM-SGAD-data a2pm_evaluation_bboxes.tar.gz --repo-type dataset --local-dir downloads
tar -xzf downloads/a2pm_evaluation_bboxes.tar.gz

For SGAD / SGAD++ training or standalone region evaluation:

hf download MrWangChi/A2PM-SGAD-data sgad_data_metadata.tar.gz --repo-type dataset --local-dir downloads
hf download MrWangChi/A2PM-SGAD-data sgad_scannet_bboxes.tar.gz --repo-type dataset --local-dir downloads
hf download MrWangChi/A2PM-SGAD-data sgad_megadepth_bboxes.tar.gz --repo-type dataset --local-dir downloads

tar -xzf downloads/sgad_data_metadata.tar.gz
tar -xzf downloads/sgad_scannet_bboxes.tar.gz
tar -xzf downloads/sgad_megadepth_bboxes.tar.gz

Preserve the MegaDepth relative symlink when extracting. Pair lists and configurations are maintained in the code repository.