DINO-SLAM: DINO-Informed Dense RGB-D SLAM

• The paper introduces a dense RGB-D SLAM framework that integrates self-supervised DINO features with a Scene Structure Encoder to capture hierarchical scene elements and structural relationships.
• It employs dual pipelines—a NeRF-based tri-plane encoding and a 3D Gaussian Splatting approach—to optimize tracking, mapping, and rendering across Replica, ScanNet, and TUM RGB-D benchmarks.
• EDINO features, derived via cross-attention between appearance and depth cues, provide continuous relational scene representations without requiring semantic annotations.

Searching arXiv for the target paper and closely related DINO-in-SLAM work to ground the article and citations. arXiv search query: "DINO-SLAM (Gong et al., 25 Jul 2025) DINO-informed RGB-D SLAM Neural Implicit Explicit Representations" DINO-SLAM is a dense RGB-D SLAM framework that introduces a DINO-informed design strategy for both neural implicit and explicit scene representations. In the formulation reported for "DINO-SLAM" (Gong et al., 25 Jul 2025), self-supervised DINO visual features are enriched by a Scene Structure Encoder (SSE) into Enhanced DINO features, denoted EDINO, with the stated goal of capturing hierarchical scene elements and structural relationships. The method is instantiated in two paradigms: a NeRF-based pipeline built on tri-plane scene encodings and a 3D Gaussian Splatting (3DGS)-based pipeline built on EDINO-augmented Gaussian parameters. On Replica, ScanNet, and TUM RGB-D, the reported results position the framework as a representation-agnostic mechanism for improving dense neural RGB-D SLAM through feature-informed tracking, mapping, and rendering (Gong et al., 25 Jul 2025).

1. Concept and problem formulation

DINO-SLAM addresses dense RGB-D simultaneous localization and mapping in settings where the system must estimate camera poses, reconstruct scene geometry, and maintain a dense scene representation suitable for rendering. The paper explicitly targets two families of neural SLAM systems: neural implicit systems based on NeRF/SDF-style volumetric rendering and neural explicit systems based on 3D Gaussian Splatting (Gong et al., 25 Jul 2025).

The stated motivation is that prior neural dense SLAM pipelines are dominated by low-level supervision, especially RGB color residuals, depth residuals, and geometric regularizers. DINO-SLAM proposes to supplement these with self-supervised DINO features and, more specifically, with EDINO features produced by the SSE. The paper attributes three advantages to a DINO-informed formulation: hierarchical scene understanding, continuous relational representation, and the absence of semantic annotations. In this framing, DINO features are not treated as discrete semantic labels but as continuous signals that can encode scene layout context, part-whole relations, and geometric-appearance coupling (Gong et al., 25 Jul 2025).

This distinguishes DINO-SLAM from systems that use DINO-family features only as an auxiliary semantic map or only in a front-end correspondence module. In DINO-SLAM, the central claim is broader: the scene representation itself should be informed by DINO-derived structure. A plausible implication is that the work is less a single architecture than a feature-integration principle spanning multiple representation families.

2. Scene Structure Encoder and EDINO

The core learned module is the Scene Structure Encoder, which takes RGB frame $c$ and depth frame $d$ and outputs raw DINO features $f_d$ together with enhanced DINO features $f_{ed}$: $\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$

The SSE begins with raw DINO extraction using ViT-small. It then derives appearance features $f_a$ from DINO features through an MLP $F_{ed}$, and geometric features $f_g$ from depth through another MLP $F_{eg}$ (Gong et al., 25 Jul 2025). The first attention stage refines DINO features using geometry and appearance: $$f_{\text{atten-d}}= \text{softmax}\left(\frac{f_g f_a^T}{\sqrt{d_a}}\right) f_d.$$ A second attention stage refines appearance using geometry and the refined DINO features: $d$0 After a further MLP $d$1, EDINO is formed by concatenation: $d$2

Within the paper’s terminology, EDINO is intended to encode hierarchical scene elements and structural relationships rather than only image-level self-supervised similarity. The architecture does not implement an explicit scene graph; instead, hierarchy is represented as a latent embedding generated by cross-interactions among DINO, appearance, and depth-derived geometry. The paper’s ablations report that raw DINO already improves performance, while the full EDINO formulation improves further, indicating that the gain is not solely attributable to adding a pretrained feature extractor (Gong et al., 25 Jul 2025).

The paper also states that DINO-SLAM does not require offline training. This suggests that the pretrained DINO backbone supplies the initial representational prior, while the SSE and downstream scene representation are optimized within the SLAM process itself. Because a separate pretraining procedure for SSE is not specified, the safest reading is that EDINO functions as an online structural prior rather than a separately supervised semantic encoder.

3. NeRF-based DINO-SLAM

The implicit formulation follows a tri-plane SDF/radiance-field style pipeline. Sampled 3D points are encoded using three tri-planes: a geometry tri-plane $d$3, an appearance tri-plane $d$4, and a feature tri-plane $d$5. Three shallow MLPs decode these into signed distance, raw color, and raw DINO feature channels: $d$6

For sampled points $d$7 on a ray, density is obtained from the SDF by

$d$8

where $d$9 is a learnable sharpness parameter. Volume-rendering weights are then

$f_d$0

Rendered DINO features, color, and depth are

$f_d$1

$f_d$2

$f_d$3

The supervision includes two feature-level losses in addition to RGB-D terms. The feature tri-plane is supervised by EDINO through

$f_d$4

and rendered raw DINO features are aligned with image DINO features through

$f_d$5

The standard RGB and depth losses are

$f_d$6

The paper further states that free space loss, signed distance loss, and truncation loss are inherited from ESLAM (Gong et al., 25 Jul 2025).

Operationally, mapping samples $f_d$7 pixels from a temporal window containing the current frame, the two most recent frames, and $f_d$8 selected keyframes. Joint optimization includes the tri-planes, decoder MLPs, and the camera poses for the selected $f_d$9 frames. For tracking, each incoming frame undergoes pose estimation, with rays lacking valid ground-truth depth or outlier pixels excluded following the base method. The paper gives the loss weights for the NeRF variant as 5.0 for the encoding feature loss and 0.01 for the DINO-feature loss, and reports tri-plane resolutions of 24 cm / 6 cm for geometry and 24 cm / 3 cm for appearance and DINO tri-planes (Gong et al., 25 Jul 2025).

4. 3DGS-based DINO-SLAM

The explicit formulation uses anisotropic Gaussian primitives $f_{ed}$0. Each Gaussian $f_{ed}$1 includes color $f_{ed}$2, radius $f_{ed}$3, opacity $f_{ed}$4, mean $f_{ed}$5, covariance $f_{ed}$6, and an EDINO feature $f_{ed}$7 inserted into a distinct feature channel. The Gaussian density is written as

$f_{ed}$8

Projection to the image plane follows

$f_{ed}$9

where $\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$0 is the projection function, $\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$1 is the camera pose, $\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$2 is the Jacobian of the projective transform approximation, and $\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$3 is the rotational part of $\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$4 (Gong et al., 25 Jul 2025).

Rendered feature, color, and depth maps are obtained by front-to-back compositing: $\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$5

$\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$6

$\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$7

The feature supervision is

$\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$8

with RGB and depth losses

$\Phi_{SSE}(c,d) \to (f_d, f_{ed}).$9

In this variant, tracking optimizes camera pose only using color and depth residuals; the paper does not state that DINO features are directly included in the tracking residual. Mapping performs global bundle adjustment to mitigate forgetting and drift, especially in large scenes and under large camera motions (Gong et al., 25 Jul 2025). Keyframe selection follows EC-SLAM’s co-visibility-based strategy together with motion thresholds inherited from Gaussian SLAM, using

$f_a$0

The paper explicitly notes that it does not introduce a new densification or pruning rule specific to EDINO, and that the implementation remains consistent with the original Gaussian Splatting framework. This suggests that EDINO primarily modifies the feature content and optimization objective of the Gaussian field rather than the life-cycle rules of Gaussian primitives.

5. Evaluation and reported performance

The evaluation covers Replica, ScanNet, and TUM RGB-D, with tracking measured by ATE RMSE, reconstruction by Accuracy, Completion, Completion rate, Depth L1, and RMSE, and rendering by PSNR, SSIM, and LPIPS (Gong et al., 25 Jul 2025).

Reported main results

Setting	Dataset	Reported result
NeRF-based DINO-SLAM	Replica	Accuracy 1.94 cm, Completion 1.64 cm, Completion rate 97.06%, Depth L1 0.66 cm, RMSE 0.51 cm
3DGS-based DINO-SLAM	Replica	PSNR 39.12, SSIM 0.971, LPIPS 0.062, RMSE 0.30 cm
NeRF-based DINO-SLAM	ScanNet	Average ATE RMSE 7.1
3DGS-based DINO-SLAM	ScanNet	Average ATE RMSE 9.5
NeRF-based DINO-SLAM	TUM RGB-D	Average ATE RMSE 1.95
3DGS-based DINO-SLAM	TUM RGB-D	Average ATE RMSE 1.36

On Replica, the NeRF variant reports the best completion, completion rate, and RMSE among the compared NeRF methods, while Point-SLAM reports better accuracy and depth L1. The paper interprets this as evidence that DINO-SLAM’s reconstruction is more complete (Gong et al., 25 Jul 2025). On the 3DGS side, the reported Replica results are the best among compared 3DGS methods on PSNR, LPIPS, and RMSE, while Gaussian-SLAM reports the highest SSIM at 0.993.

On ScanNet, the NeRF variant reports an average ATE RMSE of 7.1, behind PLGSLAM at 6.8 but ahead of ESLAM, Co-SLAM, NICE-SLAM, and Point-SLAM. The 3DGS variant reports an average ATE RMSE of 9.5, outperforming SplaTAM, MonoGS, and Gaussian-SLAM (Gong et al., 25 Jul 2025). On TUM RGB-D, the 3DGS variant reports the best average ATE RMSE at 1.36, ahead of MonoGS at 1.47, while the NeRF variant reports 1.95 and surpasses the compared NeRF baselines.

The paper also reports memory and runtime tradeoffs on Replica room0. DINO-SLAM (NeRF) is reported at 48.7 MB and 2.55 FPS, slower and larger than Co-SLAM and ESLAM but faster and lighter than Point-SLAM and SNI-SLAM. DINO-SLAM (3DGS) is reported at 27.8 MB and 0.61 FPS, essentially matching MonoGS in speed and memory while remaining much more compact than SplaTAM, SGS-SLAM, and related 3DGS baselines (Gong et al., 25 Jul 2025).

6. Ablations, interpretation, and relation to adjacent DINO-based SLAM work

The ablations are central to the paper’s argument that the contribution lies not only in importing DINO but in structurally enriching it. In the NeRF ablation on Replica, the reported RMSE is 0.63 without DINO, 0.58 without EDINO, and 0.51 for the full system. In the 3DGS ablation on ScanNet, the reported RMSE is 12.67 without DINO, 11.39 without EDINO, 13.18 without BA, and 9.55 for the full system (Gong et al., 25 Jul 2025). These values indicate that raw DINO improves over no-DINO, while EDINO improves further, and that global BA is particularly important in the explicit pipeline.

The paper also tests generality by plugging the DINO-SLAM design into multiple encodings. On Replica, the reported RMSE improves from 0.86 to 0.81 for Co-SLAM, from 0.63 to 0.51 for ESLAM, from 0.46 to 0.43 for SNI-SLAM, from 0.52 to 0.46 for Point-SLAM, from 0.36 to 0.34 for SplaTAM, and from 0.32 to 0.30 for MonoGS (Gong et al., 25 Jul 2025). This supports the representation-agnostic framing.

Within the emerging literature, DINO-SLAM occupies a different position from several contemporaneous DINO-related systems. "LEG-SLAM" (Titkov et al., 3 Jun 2025) attaches compressed DINOv2 semantic features to Gaussians for open-vocabulary querying but does not formulate DINO as the primary tracking representation. "UP-SLAM" (Zheng et al., 28 May 2025) uses DINOv2-derived signals for uncertainty estimation and dynamic masking in a Gaussian SLAM system, again without making DINO the SLAM backbone. "VGGT-SLAM++" (Mandal et al., 8 Apr 2026) uses DINOv2 embeddings of DEM tiles for retrieval, covisibility graph construction, and high-cadence local correction rather than for direct odometry. By contrast, DINO-SLAM (Gong et al., 25 Jul 2025) uses DINO-derived features to supervise the scene representation itself in both implicit and explicit pipelines.

At the same time, DINO-SLAM differs from feature-based front-end work such as "DINO-VO" (Azhari et al., 17 Jul 2025), which leverages DINOv2 for sparse feature matching in monocular visual odometry at 72 FPS with less than 1GB of memory usage but does not include mapping, loop closure, or bundle adjustment. A plausible implication is that DINO-SLAM and DINO-VO represent complementary directions: the former emphasizes dense neural representation learning in RGB-D SLAM, while the latter emphasizes geometric front-end robustness in sparse monocular VO.

The paper’s stated limitations are that the current pipelines contain only foundational components and that future work will add loop closure, sub-mapping, and more advanced system components (Gong et al., 25 Jul 2025). Additional limitations are apparent from the description: direct DINO-based pose residuals are not formulated for the 3DGS tracker, dynamic-scene behavior is not analyzed in depth, and the SSE architecture is specified at the module level rather than at a fully implementation-complete level. These absences are significant because they mark the work as a dense neural SLAM framework centered on scene representation quality rather than a complete large-scale SLAM system.

In that sense, DINO-SLAM can be understood as an attempt to establish a common principle across NeRF and 3DGS SLAM: pretrained foundation-model features become more useful for SLAM when enriched by depth-conditioned structural reasoning and then tied directly to the map representation. The reported gains on Replica, ScanNet, and TUM suggest that this principle is effective within the tested RGB-D benchmarks, while the lack of loop closure and sub-mapping indicates that system-level completeness remained outside the scope of the initial formulation (Gong et al., 25 Jul 2025).