metadata
license: mit
tags:
- kornia
- feature-detection
- feature-description
- image-matching
kornia/dedode
Pretrained weights for DeDoDe (Detect, Don't Describe — Describe, Don't Detect),
used by kornia.feature.DeDoDe.
DeDoDe trains detection and description independently: the detector is trained to find repeatable 3D points, the descriptor is trained separately for matchability. Supports upright and rotation-equivariant (C4 / SO2) variants. 3DV 2024.
Original repo: Parskatt/DeDoDe Rotation-equivariant variants: georg-bn/rotation-steerers
Weights
| File | Type | Variant |
|---|---|---|
dedode_detector_L.pth |
Detector | L-upright |
dedode_detector_C4.pth |
Detector | L-C4 |
dedode_detector_SO2.pth |
Detector | L-SO2 |
dedode_detector_L_v2.pth |
Detector | L-upright v2 |
dedode_descriptor_B.pth |
Descriptor | B-upright |
dedode_descriptor_G.pth |
Descriptor | G-upright |
B_C4_Perm_descriptor_setting_C.pth |
Descriptor | B-C4 |
B_SO2_Spread_descriptor_setting_C.pth |
Descriptor | B-SO2 |
G_C4_Perm_descriptor_setting_C.pth |
Descriptor | G-C4 |
G_SO2_Spread_descriptor_setting_C.pth |
Descriptor | G-SO2 |
Citation
@inproceedings{edstedt2024dedode,
title = {{DeDoDe}: Detect, Don't Describe --- Describe, Don't Detect
for Local Feature Matching},
author = {Johan Edstedt and Georg Bökman and Mårten Wadenbäck
and Michael Felsberg},
booktitle = {3DV},
year = {2024}
}