Long-term environmental monitoring requires the ability to reconstruct and align 3D models across repeated site visits separated by months or years. However, existing Structure-from-Motion (SfM) pipelines implicitly assume near-simultaneous image capture and limited appearance change, and therefore fail when applied to long-term monitoring scenarios such as coral reef surveys, where substantial visual and structural change is common. In this paper, we show that the primary limitation of current approaches lies in their reliance on post-hoc alignment of independently reconstructed sessions, which is insufficient under large temporal appearance change. We address this limitation by enforcing cross-session correspondences directly within a joint SfM reconstruction. Our approach combines complementary handcrafted and learned visual features to robustly establish correspondences across large temporal gaps, enabling the reconstruction of a single coherent 3D model from imagery captured years apart, where standard independent and joint SfM pipelines break down. We evaluate our method on long-term coral reef datasets exhibiting significant real-world change, and demonstrate consistent joint reconstruction across sessions in cases where existing methods fail to produce coherent reconstructions. To ensure scalability to large datasets, we further restrict expensive learned feature matching to a small set of likely cross-session image pairs identified via visual place recognition, which reduces computational cost and improves alignment robustness.
Figures & tables
Fig. 1 : Long-term multi-session 3D reconstruction for coral reef monitoring. Our method generates a single coherent 3D reconstruction from video-derived images of a coral reef captured over a three-year period. The figure presents the combined 3D point cloud alongside the fields of view of three images acquired from the same location in each year. Despite substantial temporal changes, images of the same area remain aligned with pixel-level accuracy within a shared coordinate frame.
Fig. 2 : Overview of the proposed reconstruction pipeline. RGB images acquired from multiple visits to the same area are provided to a feature extraction and matching module, together with a distance matrix generated from visual place recognition. Images within the same survey are matched exhaustively using fast handcrafted features to ensure robust intra-session reconstruction. For images across different visits, candidate cross-session pairs are first identified using the VPR-based distance matrix. A learned feature matcher is then applied selectively to these candidates to establish reliable cross-session correspondences, which are enforced directly during joint Structure-from-Motion optimization.
Fig. 3 : Multi-year AUV surveys at Sesoko Island (2016–2018). Top : approximate GPS trajectories of repeated lawnmower-pattern surveys covering the same reef area across multiple years. Bottom : example images captured at the same location before and after Typhoon Trami (Paeng), illustrating substantial appearance variation and structural disruption. These changes severely challenge cross-session correspondence establishment and long-term reconstruction.
Dataset
Year
# Images (k)
Dur. (h)
Dist. (km)
Avg. Depth (m)
Seafloor
Platform
Sesoko
2018
2.8
01:16
0.7
37
Mesophotic coral
SOTON/UTOK OPLab AUV
2017
3.0
02:04
0.7
36
2016
2.1
00:58
0.5
38
TABLE I : Technical details of the Sesoko campaign and subsampled evaluation regions. The table summarizes the full survey data collected across three years and the smaller spatial subsets used for controlled evaluation. R denotes the approximate radius of each subset region.
Fig. 4 : Examples of selected image pairs used for cross-session evaluation. The shown pairs are spatially distributed across the survey area and drawn from different years. These image pairs form the basis for manual annotation of cross-session correspondences used in the reprojection error evaluation.
Fig. 5 : Qualitative alignment of 3D points across two visits. Projections of reconstructed 3D points from two survey sessions (SSK16 in blue, SSK17 in yellow) overlaid on the images (left) and visualized in 3D (right). The top row shows the input images, manually annotated cross-session correspondences, and a warped cross-session image to indicate the expected field-of-view overlap. In the subsequent rows, yellow circles denote annotated correspondences, blue crosses indicate projected points, and red lines represent reprojection error; corresponding 3D projections are shown on the right. While COLMAP + ICP yields point clouds that appear coarsely aligned in 3D, substantial pixel-level misalignment remains. COLMAP + BUFFER-X improves global overlap but exhibits residual local inconsistencies. Our method achieves close agreement with the expected overlap, indicating accurate cross-session alignment.
Fig. 6 : Histograms of reprojection error across evaluation subsets. Each histogram shows the distribution of pixel reprojection error for manually annotated cross-session correspondences. Results for our method with and without visual place recognition (VPR) filtering are shown as part of an ablation analysis. Incorporating VPR-guided learned matching produces a tighter error distribution and lower reprojection error, indicating more accurate cross-session camera geometry and improved reconstruction consistency. Median values for each method are indicated by vertical dashed lines.
Subset
COLMAP
MapAnything
COLMAP + ICP
COLMAP + BUFFER-X
Ours
ssk-s01
✗
✗
162.94
27.07
2.25
ssk-s02
✗
✗
76.94
23.19
2.69
ssk-s03
✗
✗
121.83
51.46
3.81
ssk-s04
✗
✗
299.34
206.72
5.85
Avg. RPE
✗
✗
165.26
77.11
3.65
TABLE II : Comparison of alignment methods on different subsets using median reprojection error (pixels). Lower values indicate more accurate alignment across sessions, highlighting the substantial improvement achieved by our method. The best result for each subset is in bold ; cross marks (✗) indicate results where the method failed to produce a complete point cloud incorporating points from all sessions.
Fig. 7 : Qualitative comparison of joint reconstruction strategies under long-term appearance change. (a) Joint reconstruction using COLMAP with exhaustive handcrafted feature matching fails to incorporate all visits, resulting in incomplete reconstructions. (b) MapAnything produces fragmented and misaligned geometry across visits. (c, d) COLMAP with post-hoc alignment techniques produces point clouds that appear coarsely aligned, but result in discrepancies at the pixel-level (see Figure 5 ). (e) Our approach successfully integrates imagery from all visits into a single coherent reconstruction by enforcing cross-session correspondences during reconstruction.
Fig. 8 : Reprojection error for 3D points observed within single visits and across multiple visits. Points observed across multiple survey visits exhibit higher reprojection error due to increased uncertainty under long-term appearance and structural change. This behavior is expected and highlights the increased difficulty of cross-session reconstruction, motivating the need for explicit enforcement of cross-session correspondences.
Fig. 9 : Feature matching time across increasing dataset sizes for four matching strategies with different levels of selectivity. Constraining learned feature matching to cross-session image pairs, and further guiding candidate selection using visual place recognition (VPR), substantially improves scalability compared to exhaustive approaches. As shown in Figure 6 , VPR-guided matching not only reduces computational cost but also yields lower and more concentrated reprojection error distributions.
Establishing consistent correspondences across images is essential for 3D vision tasks such as structure-from-motion (SfM), yet most existing matchers operate in a pairwise manner, often producing fragmented and geometrically inconsistent tracks when their predictions are chained across views. We propose MV-RoMa, a multi-view dense matching model that jointly estimates dense correspondences from a source image to multiple co-visible targets. Specifically, we design an efficient model architecture which avoids high computational cost of full cross-attention for multi-view feature interaction: (i) multi-view encoder that leverages pair-wise matching results as a geometric prior, and (ii) multi-view matching refiner that refines correspondences using pixel-wise attention. Additionally, we propose a post-processing strategy that integrates our model's consistent multi-view correspondences as high-quality tracks for SfM. Across diverse and challenging benchmarks, MV-RoMa produces more reliable correspondences and substantially denser, more accurate 3D reconstructions than existing sparse and dense matching methods. Project page: https://icetea-cv.github.io/mv-roma/.
Online monocular 3D reconstruction enables dense scene recovery from streaming video but remains fundamentally limited by the stability-adaptation dilemma: the reconstruction model must rapidly incorporate novel viewpoints while preserving previously accumulated scene structure. Existing streaming approaches rely on uniform or attention-based update mechanisms that often fail to account for abrupt viewpoint transitions, leading to trajectory drift and geometric inconsistencies over long sequences. We introduce PAS3R, a pose-adaptive streaming reconstruction framework that dynamically modulates state updates according to camera motion and scene structure. Our key insight is that frames contributing significant geometric novelty should exert stronger influence on the reconstruction state, while frames with minor viewpoint variation should prioritize preserving historical context. PAS3R operationalizes this principle through a motion-aware update mechanism that jointly leverages inter-frame pose variation and image frequency cues to estimate frame importance. To further stabilize long-horizon reconstruction, we introduce trajectory-consistent training objectives that incorporate relative pose constraints and acceleration regularization. A lightweight online stabilization module further suppresses high-frequency trajectory jitter and geometric artifacts without increasing memory consumption. Extensive experiments across multiple benchmarks demonstrate that PAS3R significantly improves trajectory accuracy, depth estimation, and point cloud reconstruction quality in long video sequences while maintaining competitive performance on shorter sequences.
Lanbo Xu, Liang Guo, Caigui Jiang +1
State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, China · University of East Anglia, Norwich, NR47TJ, United Kingdom
We present LIST3R, an instance-aware framework for long-sequence 3D reconstruction inspired by the way humans organize spatial memory around stable and recognizable objects. LIST3R organizes long-sequence reconstruction around instance anchors, using them to reconnect fragmented subsequences and consolidate local observations into a coherent global 3D scene. Given a long video, our approach partitions it into overlapping subsequences and builds a structured local instance library for each partial reconstruction, maintaining persistent trackable anchors with semantic and geometric evidence. These anchors are matched across subsequences to recover revisited regions and provide object-aware constraints for fragment alignment, producing a consistent global reconstruction. During this process, the evolving geometric evidence updates the local instance libraries and progressively organizes them into a unified global 3D instance library. Experiments on long-sequence benchmarks show that our method produces more accurate trajectories and higher-quality 3D reconstructions, highlighting the effectiveness of persistent instance anchors for organizing long-horizon 3D reconstruction. Our code is available on the project page: https://yixn965.github.io/LIST3R/.
Jing Gao, Wei Wang, Feiran Wang +1
Beijing Jiaotong University · University of Illinois Chicago