[ECCV '24 Workshop Baseline]: Matching 2D Images in 3D: Metric Relative Pose from Metric Correspondences, without using multiple images or device poses.
Only using the last frame in the query sequence, not using device poses.

[Baseline] MicKey (using only a single frame)

Map-Free Visual Relocalization Enhanced by Instance Knowledge and Depth Knowledge use only single frame

Map-Free Visual Relocalization Enhanced by Instance Knowledge and Depth Knowledge

A MASt3R decoder fine-tuned over DUNE, a ViT-Base Distilled UNiversal Encoder [CVPR 2025]

DUNE + MAST3R

Forward with (ref, query) gives pointmaps and pixel matches. These latter ones are used for Ess. matrix estimation, and the scale is obtained from the predicted pointmaps.
(re-submission of previous results)

MASt3R (Ess.Mat + D.Scale)

Metric3d for 3d scaled depth estimation and loftr for initial patch matching.Then an geometric pnp algorithm is used for pose estimation.

interp_metric3d_loftr_3d2d

[ECCV '24 Workshop Baseline]: RoMa correspondences with MicKey depth maps predictions.

RoMa w/ MicKey depth maps

Matching 2D Images in 3D: Metric Relative Pose from Metric Correspondences

MicKey trained w/ Overlap Score

MicKey

[ECCV '24 Workshop Baseline]: SuperPoint + SuperGlue correspondences with MicKey depth maps predictions.

SuperGlue w/ MicKey depth maps

[ECCV '24 Workshop Baseline]: LoFTR correspondences with MicKey depth maps predictions.

LoFTR w/ MicKey depth maps

Using the depth map from the Depth Anything Model as the ground truth, we trained the Mickey Variant network in a supervised manner.

Mickey Variant GT_Depth

Our flexible method produces accurate and robust pose estimates using complementary strengths of Correspondence + Solver and Learning-Based methods.

FAR (LoFTR)

DUSt3R --> pixel correspondences between output 3d points

DPT-KITTI & DUSt3R (Ess.Mat. + D.Scale)

FAR (SuperGlue)

Same as loftr one, 2px opencv MAGSAC etc.

DPT-KITTI & RoMa (Ess.Mat. + D.Scale) Note: maybe buggy

We replace the baseline DPT-KITTI with KBR++ depth predictions. KBR++ is a self-supervised monocular depth estimation system trained on a combination of SlowTV, CribsTV, MannequinChallenge & Kitti. The model is transferred to MapFreeReloc in a zero-shot fashion, without fine-tuning. 

Since KBR++ is trained only on monocular data (ie without a stereo baseline), the predictions are scaleless. Metric-scale predictions are obtained by aligning KBR++ to the provided baseline DPT-KITTI using least-squares.

KBR++ & LoFTR (PnP)

We replace the baseline DPT-KITTI with KBR depth predictions. KBR is a self-supervised monocular depth estimation system trained on a combination of SlowTV, MannequinChallenge & Kitti. The model is transferred to MapFreeReloc in a zero-shot fashion, without fine-tuning. 

Since KBR is trained only on monocular data (ie without a stereo baseline), the predictions are scaleless. Metric-scale predictions are obtained by aligning KBR to the provided baseline DPT-KITTI using least-squares.

KBR & LoFTR (PnP)

[ECCV'24 Workshop Baseline]: ASpanFormer: Detector-Free Image Matching
with Adaptive Span Transformer

DPT-KITTI & ASpanFormer (Ess.Mat. + D.Scale)

DPT-KITTI & LoFTR (PnP)

DPT-KITTI & LoFTR (Ess.Mat. + D.Scale)

DPT-KITTI & SuperGlue (Ess.Mat. + D.Scale)

DPT-KITTI & SuperGlue (PnP)

[ECCV '24 Workshop Baseline]: DISK+LightGlue correspondences with DPT-KITTI depth maps.

DPT-KITTI & LightGlue w/ DISK (Ess.Mat. + D.Scale)

[ECCV '24 Workshop Baseline]: DISK correspondences with DPT-KITTI depth maps.

DPT-KITTI & DISK (Ess.Mat. + D.Scale)

[ECCV'24 Workshop Baseline]: DeDoDe: Detect, Don't Describe -- Describe, Don't Detect for Local Feature Matching

Evaluation Leaderboard: Single Frame