REMIND: RE-identification with Memory for INDoor Navigation

Pablo Diaz-Pereda, Alejandro Rodriguez-Ramos, David Perez-Saura, and Pascual Campoy

Computer Vision and Aerial Robotics Group at Centre for Automation and Robotics C.A.R. (UPM-CSIC), Universidad Politecnica de Madrid (CVAR-UPM), Calle Jose Gutierrez Abascal 2, 28006 Madrid, Spain

Paper Method Code

REMIND overview video.

To read the visualization, follow each object's color and class label over time: when they remain stable, REMIND is preserving the same identity for the same physical object. White masks indicate ambiguous or provisional tracks, where the system delays committing to a persistent identity until there is enough evidence.

Abstract

Mobile robots operating indoors must re-identify previously observed objects after long temporal gaps, significant viewpoint changes, and severe illumination variations. This remains a challenging problem: multi-object tracking methods are optimized for short-term association of pedestrians and vehicles at video rates, person and vehicle re-identification approaches lack persistent memory mechanisms, and state-of-the-art video object segmentation techniques rely on reactive distractor filtering rather than enforcing global identity consistency.

To address these limitations, we present REMIND, an online tracker designed for long-term multi-object re-identification of generic indoor objects from monocular RGB imagery, requiring neither camera pose nor depth. Motivated by evidence from visual cognition that humans rely on accumulated appearance familiarity and spatial context rather than explicit self-localization, REMIND combines frozen DINOv3 features with a dual-bank multi-prototype appearance memory, part- and background-level descriptors, a neighbour-context reasoning module exploiting spatial co-occurrence, and joint Hungarian assignment with ambiguity-aware safeguards. On a purpose-built indoor dataset featuring controlled revisits and dense same-class clutter, REMIND reaches 90.35% IDF1, nearly 20 points above a state-of-the-art video object segmentation baseline and more than 36 above a strong tracking-by-detection baseline. On ScanNet++, it attains the highest IDF1 in every setting but one, end-to-end detection over all scenes, where the tracking-by-detection baseline is marginally ahead while REMIND still associates and recovers identities more accurately; it also completes every scene, whereas the video object segmentation baseline exhausts GPU memory on 66.9% under YOLO detections. The complete system, evaluation framework, and dataset are publicly released.

Datasets
Custom Dataset Single video recorded in a living room under controlled conditions. The sequence contains medium-low object density and a medium level of repeated object classes, with controlled viewpoint and illumination variation.
Custom dataset GT masks for frame 000045 Custom dataset GT masks for frame 000105 Custom dataset GT masks for frame 000187 Custom dataset GT masks for frame 000278 Custom dataset GT masks for frame 000816 Custom dataset GT masks for frame 000930
ScanNet++ Indoor scenes with object instances, semantic labels, and segmentation masks. We convert selected scenes to a DAVIS-style tracking format to evaluate identity consistency across complex real-world indoor environments. The GIFs shown below use the official ScanNet++ benchmark annotations from the official dataset release.
ScanNet++ scene f1e01af60a with benchmark GT masks and class labels ScanNet++ scene 3ff873c77e with benchmark GT masks and class labels ScanNet++ scene d3b7d054c4 with benchmark GT masks and class labels ScanNet++ scene 07e6f56969 with benchmark GT masks and class labels ScanNet++ scene 41b00feddb with benchmark GT masks and class labels ScanNet++ scene 45b0dac5e3 with benchmark GT masks and class labels
ScanNet++ page Code
Tracking Examples

These examples show REMIND tracking on selected ScanNet++ indoor scenes converted to DAVIS-style sequences.

Main Results

REMIND explicitly models ambiguous and uncertain cases. For quantitative evaluation, ambiguous/provisional labels are collapsed into a final identity decision so standard tracking metrics can be computed. This collapsed output is used only for reporting: internally, the method does not always force an immediate identity label.

MASA is evaluated with an adapted MASA-GDINO setup: class-aware association, full-scene memory, disabled spatial distance filtering, and 0.3 embedding momentum. ScanNet++ rows report the DAM4SAM-comparable subset; the YOLO subset uses the V2 forced split.

Evaluation with ground-truth masks

Custom Dataset

Method Detector IDF1 Recovery
DAM4SAM GT masks init 70.53% 69.15%
MASA GT masks 53.84% 45.74%
REMIND GT masks 90.35% 85.46%

ScanNet++

Method Detector IDF1 Recovery
DAM4SAM GT masks init 60.71% 41.65%
MASA GT masks 57.01% 42.65%
REMIND GT masks 66.67% 58.36%

Evaluation with YOLO detections

Custom Dataset

Method Detector IDF1 Recovery
DAM4SAM YOLO init 64.54% 65.96%
MASA YOLO 45.76% 38.65%
REMIND YOLO 75.45% 79.08%

ScanNet++

Method Detector IDF1 Recovery
DAM4SAM YOLO init 57.97% 53.59%
MASA YOLO 60.38% 46.89%
REMIND YOLO 61.30% 46.23%
Qualitative Results

Qualitative comparisons focus on the effect of the detection source in the custom video and on the comparison among REMIND, DAM4SAM, and MASA in selected ScanNet++ benchmark scenes.

In these videos, stable colors and class labels are the main cue for correct re-identification: they show that the tracker keeps assigning the same identity to the same object across frames. White masks mark ambiguous or temporary assignments, reflecting cases where REMIND is intentionally cautious before updating its memory.

Custom video: GT masks vs YOLO detections

Side-by-side REMIND comparison on the custom video, contrasting tracking from ground-truth masks with tracking from YOLO detections.

YOLO is expected to perform worse than ground-truth masks because detections can fail, mask boundaries are less precise, and a single object may be split into multiple regions. These fragments create false duplicate objects and can pollute the contextual memory used to describe nearby objects.

Detector: GT masks Detector: YOLO

REMIND vs DAM4SAM vs MASA on ScanNet++

Three-way qualitative comparisons on five official ScanNet++ benchmark scenes converted to DAVIS-style tracking sequences.

In the first ScanNet++ example, REMIND keeps the bed region and the curtain more consistently tracked, while DAM4SAM intermittently loses them. MASA struggles especially in the bed area, where it confuses the blankets and pillows, while the other methods preserve these identities more reliably.

REMIND DAM4SAM MASA

In the second example, DAM4SAM loses the wardrobe, the blanket, and one of the curtains, while REMIND keeps these objects more consistently tracked. MASA also struggles with the pillows, wardrobe, and curtain; although DAM4SAM exhibits related failures in this area, REMIND handles these objects more reliably than both baselines.

REMIND DAM4SAM MASA

In the third example, DAM4SAM loses the table and several objects placed on top of it at some point in the sequence, while REMIND preserves more stable identities. MASA behaves similarly to DAM4SAM: after the temporal jump, it recovers several objects with incorrect identities.

REMIND DAM4SAM MASA

In the fourth example, DAM4SAM temporarily loses the table, chair, and TV, although it later recovers them, whereas REMIND keeps these objects more consistently connected across the sequence. MASA shows behavior similar to DAM4SAM, with comparable instability in maintaining these identities.

REMIND DAM4SAM MASA

In the fifth example, REMIND maintains stable tracks for the chairs, table, and surrounding furniture, while DAM4SAM loses the chairs and table and also struggles with the door and cabinet during parts of the sequence. MASA likewise becomes confused in some frames, particularly around the chairs and shelving unit.

REMIND DAM4SAM MASA

Acknowledgements

This work has been supported by the project SHEREC "Safe Healthy and Environmental Ship Recycling", Reference: 101136056, funded by the European Union under the Horizon Europe Program HORIZON-CL4-2023-HUMAN-01 CNECT.

Citation

Please cite REMIND, MASA, and the ScanNet++ dataset when using this project page, code, or data-derived results.

REMIND

@misc{remind2026,
  title  = {REMIND: RE-identification with Memory for INDoor Navigation},
  author = {Diaz-Pereda, Pablo and Rodriguez-Ramos, Alejandro and Perez-Saura, David and Campoy, Pascual},
  year   = {2026}
}

MASA

@inproceedings{li2024masa,
  title     = {Matching Anything by Segmenting Anything},
  author    = {Li, Siyuan and Ke, Lei and Danelljan, Martin and Piccinelli, Luigi and Segu, Mattia and Van Gool, Luc and Yu, Fisher},
  booktitle = {IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2024}
}

ScanNet++

@inproceedings{yeshwanthliu2023scannetpp,
  title={ScanNet++: A High-Fidelity Dataset of 3D Indoor Scenes},
  author={Yeshwanth, Chandan and Liu, Yueh-Cheng and Nie{\ss}ner, Matthias and Dai, Angela},
  booktitle = {Proceedings of the International Conference on Computer Vision ({ICCV})},
  year={2023}
}