A benchmark for the evaluation of RGB-D SLAM systems
1. Paper overview
In one sentence: Independent camera-pose measurements make RGB-D trajectory comparisons reproducible, provided calibration, temporal association and the chosen error metric remain explicit. e01e02e04e07e09e10e11e12e13e17
| At a glance | What to know |
|---|---|
| Research problem | Source description RGB-D SLAM needed shared sequences with independent ground truth and consistent end-to-end evaluation. Small internal optimization residuals alone do not establish an accurate trajectory. Comparing estimated camera motion with calibrated external measurements makes drift and global consistency inspectable, while leaving map utility as a separate question. e02e04e11 |
| Core mechanism | Source description The paper reports 39 indoor sequences, with handheld and Pioneer 3 recordings, color/depth at 640 × 480 and 30 Hz, and motion-capture ground truth at 100 Hz. It states that the data are released under CC-BY 3.0. e02 |
| A key reported result | Ground-truth-to-camera extrinsic calibration: 3.25 mm for fr1; 4.03 mm for fr2. Mean corner-position residual; lower is better.. Calibration checkerboard corners observed in RGB versus corners predicted from motion capture, separately for fr1 and fr2. These are two calibrated sensor/environment configurations, not competing SLAM algorithms. Measured residuals combine motion-capture and visual detector noise. They are not direct measurements of trajectory ATE. e09 |
| Reading caution | Author claim The authors infer ground-truth errors below 1 mm/0.5° relatively and 10 mm/0.5° absolutely, and restrict valid comparisons to system errors significantly above those levels. These are author conclusions from calibration, not independent guarantees. e09 |
Core contributions
- Source description
The paper reports 39 indoor sequences, with handheld and Pioneer 3 recordings, color/depth at 640 × 480 and 30 Hz, and motion-capture ground truth at 100 Hz. It states that the data are released under CC-BY 3.0. e02
Figure 1. Two indoor environments and two camera mounts share an external pose-measurement strategy. Original paper, p. 1 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the top row as the environments and the bottom row as the recording platforms. The fr1 office contains desks and nearby objects; the fr2 hall offers a larger and more sparsely textured setting. In the lower panels, the rods ending in reflective markers identify the rigid body tracked by external cameras. These markers do not directly specify the color camera's optical center: Section VI-C supplies the extrinsic calibration linking those frames. The recorded images enter a SLAM system, while the independently measured, transformed poses provide its evaluation reference. e02e05e06e09e11
What it supports. The benchmark's key mechanism is the combination of RGB-D observations with an external camera-pose reference. Handheld and robot-mounted recordings broaden the motion settings while preserving this measurement principle. The photographs establish the physical arrangement; the later calibration measurements establish its reported precision.
Where the evidence stops. The robot is a recording platform, and Section III says it was manually driven. These photographs provide no evidence of autonomous planning, learned action prediction, or successful execution of a navigation policy.
2. Motivation
2.1 The problem and the proposed response
RGB-D SLAM needed shared sequences with independent ground truth and consistent end-to-end evaluation. Small internal optimization residuals alone do not establish an accurate trajectory. Comparing estimated camera motion with calibrated external measurements makes drift and global consistency inspectable, while leaving map utility as a separate question. e02e04e11
2.2 What this reading follows
A reconstructed room may look convincing while its camera trajectory drifts, and a small internal optimization residual may conceal a poor map. Sturm and colleagues address this evaluation problem with Kinect recordings paired with motion-capture poses and two trajectory metrics. Read the paper as a chain of measurements: images and reflective markers establish observations, spatial and temporal calibration connect them, and RPE or ATE turns estimated motion into an error report. Its evidence is strongest for this measurement pipeline. The illustrative SLAM comparisons help explain the metrics, but do not form a comprehensive algorithm leaderboard. e01e02e04e07e09e10e11e12e13e17
3. Research context
We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.
| Catalog dimension | Recorded classification |
|---|---|
| Major category | Datasets |
| Architecture | Not applicable |
| Prediction paradigm | Not applicable |
| Quadrant | Not applicable |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded dataset, RGB-D scene, localization and evaluation categories fit the actual contribution. Architecture, prediction paradigm and quadrant are appropriately not applicable: independent pose measurements and trajectory metrics do not constitute a One Model architecture, joint future/action prediction or inverse dynamics. e02e11e17
This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.
4. Problem formulation
4.1 Inputs and outputs
| Inputs | Outputs |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Turn a tracked marker rig into a camera-pose reference
The motion-capture system sees reflective markers, while the SLAM system sees RGB-D images. Those are different measurement frames, so a marker trajectory cannot simply be treated as the camera's trajectory. The paper first calibrates the external cameras, estimates Kinect intrinsics and checks depth scale. A marked checkerboard then relates the tracked rigid body to the Kinect optical frame. Comparing visually detected board corners against corners predicted from motion capture validates that transformation, although the residual combines errors from both measurement paths. Temporal calibration adds another necessary connection: the authors search for the offset giving the smallest board residual. Finally, the evaluator uses poses of the RGB optical frame, matching the frame into which the depth images have been registered. e06e07e08e09e10e11
5.2 Choose whether the question is local drift or global consistency
For RPE, begin with two instants separated by Δ. Compute the motion between them in the estimated trajectory and in ground truth, then compare those relative transformations. Repeating the operation yields local discrepancies whose translational components are summarized by RMSE. The interval is part of the evaluation definition: the paper distinguishes one frame from thirty frames at 30 Hz. ATE asks a different question. Because the estimated trajectory may start in an arbitrary reference frame, it first fits a single rigid transformation to ground truth, then measures residual position differences at associated instants. This alignment makes the overlaid trajectory in Figure 7 meaningful. Averaging relative errors over multiple intervals offers another global view, but requires an explicit valid-pair and sampling convention. e11e12e13e15
5.3 Interpret the score inside the measurement's limits
Reader analysis: a trajectory metric inherits the quality and coverage of its reference. The authors' calibration-based accuracy bounds imply that a small measured difference between systems may approach the uncertainty of the benchmark itself. Missing reference poses on the large tours create a different problem: scoring the observed endpoints does not reveal what happened everywhere between them. Neither problem is solved by averaging more reported digits. Map quality is a further boundary, explicitly acknowledged by the paper; an apparently accurate path can coexist with a map defect that blocks a doorway. Finally, the RPE/ATE scatter is evidence about the authors' estimated trajectories. It should motivate checking metric sensitivity on a new motion regime, rather than assuming that two translational summaries always rank systems identically. e05e09e11e12e13e15
5.4 Training and inference
During training
This is a dataset and metric paper: “training” denotes the available development partition, not an optimization objective, learned representation or specified neural training schedule. e04e17
During inference
A submitted SLAM or odometry system processes the RGB-D sequence and outputs camera poses. The evaluator scores those poses against ground truth. The paper specifies neither an action-extraction stage nor an autonomous feedback controller; robot trajectories were collected with a joystick. e05e11
5.5 Implementation flow
- Choose a sequence and protocol
The paper separates offline training sequences from website-only test evaluation to limit overfitting. Section III identifies the 19 sequences in Table I as training sequences. Debugging motions approximately isolate translation or rotation; the two large tours have ground truth only at their beginning and end. e04e05
- Acquire and register observations
Two Kinect units serve fr1 and fr2. OpenNI projects depth into 3D and back into the color view with a Z-buffer and factory calibration. Eight external cameras triangulate passive markers; extrinsic calibration transfers the tracked rigid-body pose to the RGB optical frame. e06e09e11
- Calibrate geometry and timing
Checkerboards determine camera intrinsics, depth scale and the marker-to-camera transform. The distributed depth already includes scale correction. Motion-capture/color timing is corrected, but original RGB/depth timestamps remain, requiring explicit association choices. e07e09e10
- Evaluate associated trajectories
Real streams require association and interpolation because rates, lengths and missing samples differ. Fixed-interval RPE compares relative motions; ATE first fits a rigid transformation and then compares corresponding positions. Lower translational RMSE is better for both. e11e12e13
6. Experiments & results
This benchmark pairs Kinect color and registered depth images with externally measured camera poses, then evaluates estimated trajectories through local relative pose error (RPE) and globally aligned absolute trajectory error (ATE). Its contribution is calibrated data and a comparison protocol. The strongest quantitative evidence concerns sensor and ground-truth calibration; algorithm comparisons are illustrative plots. Robot-mounted recordings were manually driven and do not demonstrate autonomous control.
This benchmark supplies physical capture photographs rather than a model architecture diagram. Its two tables contain dataset statistics and intrinsic calibration, not aggregate algorithm performance. The edition therefore uses Table I as the quantitative dataset table and Figures 6–7 as the algorithm-result visuals. There is no controlled component ablation; depth and timing calibration diagnostics in Figures 4–5 serve that explanatory role without being presented as ablation experiments. e02e07e10e14e15e16e17
6.1 Read the original evidence
Table I. The listed sequences differ substantially in duration and motion, even within a diagnostic category. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the group headings, then compare the three numerical columns: duration in seconds, average translational velocity in metres per second, and average rotational velocity in degrees per second. The fr1/rpy row reports 50.15 degrees per second, whereas fr2/rpy reports 5.77; similar sequence names therefore do not imply matched motion difficulty. Section III identifies these 19 entries as training sequences. The local heading “Testing and Debugging” describes a diagnostic sequence category, not the held-out test partition discussed in Section II. The table contains dataset measurements, not SLAM accuracy scores. e04e05
What it supports. A comparison should identify exact sequences and their motion conditions. For example, fr1/desk lasts 23 seconds with average translation of 0.41 m/s, while fr2/desk lasts 99 seconds at 0.19 m/s. Pooling them into an unspecified “desk” result would hide important differences in exposure to drift and motion.
Where the evidence stops. Average velocities do not describe brief accelerations or tracking outages. The table also does not show ground-truth coverage: Section III says the two large tours have reference poses only near the start and finish.
Figure 6. A local-error trace reveals typical drift and isolated failures that an aggregate alone would hide. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow time along the horizontal axis and translational relative pose error in metres along the vertical axis. The blue solid curve is RBM; the red dashed curve is GICP. Compare both the low-error portions and the intermittent spikes rather than treating either method as uniformly better at every instant. Section VII-A defines each error from a pair of poses separated by an interval Δ, so this is a relative-motion diagnostic. The original plot's upper boundary is 0.10 m, and some spikes reach that boundary; their full heights cannot be recovered from this figure. e12e14
What it supports. The authors describe RBM as having lower drift and fewer outliers on this illustrated fr1/desk comparison. The visible trace supports a qualitative comparison while also showing a substantial RBM outlier. It demonstrates how the metric can expose local failures, without supplying an exact aggregate improvement.
Where the evidence stops. The paper does not state the exact Δ for this plot or report its aggregate RMSE and uncertainty. Algorithm details are referred to another paper. The figure therefore cannot support a numerical benchmark-wide performance claim.
Figure 7. Global alignment makes positional error visible; the scatter compares two summaries on the authors' estimates. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Panel (a) places the true trajectory in black and the estimated trajectory in blue, with red connectors showing positional disagreement after alignment. Its x and y axes are in metres; it is a projection of the trajectory, not a reconstructed map. Panel (b) puts ATE on the horizontal axis and RPE on the vertical axis, both on logarithmic scales. The red diagonal denotes equal numerical values, not a fitted regression. Blue crosses above it have larger RPE than ATE in these examples. Check the construction against Equations (4)–(5): ATE uses a least-squares rigid alignment before extracting translation. e12e13e15e11
What it supports. The figure connects a spatially interpretable residual with a comparison of trajectory summaries. The authors' RGB-D SLAM estimates show strongly related RPE and ATE values across the illustrated range. This supports their practical observation that the metrics often produced similar orderings in their experiments.
Where the evidence stops. The accompanying prose asserts that RPE is always at least ATE, but the plotted examples do not prove a universal inequality. Both RMSE definitions extract translation; orientation-only errors need separate scrutiny, especially when the camera does not translate.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Ground-truth-to-camera extrinsic calibration Calibration checkerboard corners observed in RGB versus corners predicted from motion capture, separately for fr1 and fr2. | 3.25 mm for fr1; 4.03 mm for fr2. Mean corner-position residual; lower is better. | These are two calibrated sensor/environment configurations, not competing SLAM algorithms. Measured residuals combine motion-capture and visual detector noise. They are not direct measurements of trajectory ATE. e09 |
| Kinect depth-scale calibration Checkerboard-based calibration of the two Kinect units. | fr1: 1.035; fr2: 1.031. Multiplicative depth correction factor ds. | Unity would leave the original depth scale unchanged. These are fitted calibration parameters, already applied to released depth images; they are not accuracy scores. e07 |
| Motion-capture/color synchronization Checkerboard residual evaluated over candidate time offsets. | Approximately 20 ms. Estimated temporal lead of motion-capture poses over Kinect color images. | Figure 5b shows a residual minimum near this offset. The paper says this delay is already corrected. It is distinct from the separately observed RGB/depth arrival offset. e10 |
| Visual odometry drift on fr1/desk RBM and GICP trajectories illustrated in Figure 6; further algorithm details are delegated to reference [44]. | Qualitative plotted result: the authors report lower drift and fewer outliers for RBM. Translational relative pose error versus time, in metres. | GICP is the comparison curve; RBM also has a prominent outlier. No numerical aggregate RMSE, confidence interval or exact Figure 6 interval Δ is supplied here. The plot does not establish a benchmark-wide ranking. e14 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 4. Depth scale bias and distance-dependent noise are separate calibration effects. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. In panel (a), the horizontal axis is measured distance and the vertical axis is checkerboard-derived depth. The black diagonal denotes ideal agreement; red fr1 and blue fr2 observations reveal departures from that relationship. Panel (b) changes the question: its horizontal axis is checkerboard distance, and its vertical axis is the standard deviation of depth error, both in metres. The red solid and blue dotted curves rise as the board moves farther away. Cross-check this distinction with Section VI-B: a fitted scale correction addresses systematic bias, whereas the residual-noise experiment examines variation after calibration. e07e16
What it supports. A depth value can require scale correction and still remain noisy after that correction. Both sensors show increasing variation with distance, so a calibration factor alone cannot establish equal geometric reliability for nearby and distant surfaces. The paper states that its released depth images already include the fitted scale correction.
Where the evidence stops. Section VI-B approximates noise as 5 cm at 4 m, while the plotted curves there appear lower. Preserve that prose approximation as such; the graph supports the rising trend, not an exact 5 cm reading at 4 m.
Figure 5. The calibration target exposes a timing offset through the minimum of a geometric residual. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The checkerboard in panel (a) can be observed both visually and through its reflective markers. Panel (b) varies the assumed temporal offset and evaluates the disagreement between those observations; read the horizontal axis in milliseconds and the residual axis in millimetres. The red curve reaches its minimum near the marked 20 ms position. Section VI-D resolves the physical direction: motion-capture poses were earlier than the Kinect color images. This is a calibration residual curve, not a SLAM trajectory score, and the plotted offset convention should not be turned into an unverified timestamp-editing command. e09e10
What it supports. Temporal registration is part of obtaining a trustworthy pose reference. The paper reports an approximately 20 ms motion-capture lead and says that this correction is already applied to the dataset. The remaining user-facing association problem concerns the separate color and depth image timestamps.
Where the evidence stops. Do not apply the motion-capture correction twice. Color/depth arrival timing is a different offset whose timestamps were intentionally retained. The curve supplies neither an uncertainty interval nor a universal offset for other hardware.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors infer ground-truth errors below 1 mm/0.5° relatively and 10 mm/0.5° absolutely, and restrict valid comparisons to system errors significantly above those levels. These are author conclusions from calibration, not independent guarantees. e09
Depth noise grows with distance; rolling shutter and changing exposure affect moving-camera images. Large tours lose intermediate ground truth. Accurate trajectories do not ensure accurate maps or navigable free space. e05e16e11
Figure 7 supports correlated RPE/ATE values for the illustrated RGB-D SLAM estimates, not universal metric equivalence. Both displayed RMSE formulas extract translation; rotational fidelity merits separate inspection. e12e13e15
7.2 Questions for discussion
- How should evaluation report failures and missing ground truth without rewarding short surviving trajectories?
- When would translational ATE/RPE conceal orientation errors relevant to reconstruction?
8. Reproducibility audit
8.1 Requirements and known gaps
For dataset replay, decode 16-bit depth as metres scaled by 5000, retain timestamp associations and export translation plus unit quaternion in the RGB optical frame. The acquisition stack is Ubuntu 10.10, ROS Diamondback and OpenNI; accelerometer recording is 500 Hz. e06e11
Specify association tolerance, interpolation, evaluated coverage, RPE interval and any relative-pair sampling. The paper proposes sampled approximation to quadratic all-interval evaluation but does not fix these experiment settings. Equation (3) includes Δ = n although m = n − Δ then vanishes; a literal implementation needs an explicit valid-pair boundary. e10e11e12
8.2 Proposed reproduction checks
The following checks are proposals motivated by the paper. They have not been run as part of this reading.
Check 1: Test the corrected timing convention without changing frame coverage
Reader-proposed check, not performed: on the documented checkerboard calibration recordings, compare visual corners with motion-capture-predicted corners while sweeping an additional offset from −40 to +40 ms. Use the same valid frame subset, extrinsic transform, detector and interpolation rule at every offset. For the already corrected release, test whether the residual minimum lies near zero; then inject a known 20 ms shift as a positive control and check that the recovered minimum moves by the corresponding amount. Report the residual curve and uncertainty of the minimum, not only its best point. A persistent nonzero optimum would challenge the assumed timing convention or reveal a remaining calibration interaction. e05e09e10e11
Check 2: Separate coordinate invariance from sensitivity to orientation errors
Reader-proposed check, not performed: evaluate exact pose sequences under three controlled changes. First, apply one rigid global coordinate change to an otherwise exact estimated trajectory; associated RPE and rigidly aligned ATE should remain zero within numerical precision. Second, hold all positions at one point while introducing time-varying orientation error; compare the two translational summaries with an explicitly reported rotational error. Third, add position perturbations as a positive control for the translational metrics. Keep timestamps, valid pairs and RPE intervals fixed, and exclude the zero-pair Δ = n endpoint. Nonzero angular error with zero translational scores would demonstrate why Figure 7's practical correlation cannot substitute for checking rotational fidelity. e11e12e13e15
8.3 Reading coverage
Visual audit: All eight original PDF pages were rendered and visually inspected, including title/author affiliations, Figures 1–7, Tables I–II, calibration details, timing conventions, Equations (1)–(5), acquisition software/hardware and references. All six final original crops were individually inspected after margin corrections. Figure 4's scale/noise axes and Figure 5's timing direction were checked against their captions and Sections VI-B/VI-D. Figure 7's diagonal was checked against the metric equations and accompanying claim. The original Figure 6 truncates some spike heights at its plotted upper limit. No appendix is present. External code, dataset archives, videos and separate supplements remain outside this review.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. Appendix coverage: not present.
Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.
Text reading scope & known omissions
- PDF p. 1: title, author affiliations, abstract and I. Introduction
- PDF pp. 1–2: II. Related Work
- PDF pp. 2–4: III. Dataset, including calibration, debugging, handheld and robot sequences
- PDF p. 4: IV. Data Acquisition; V. File Formats, Tools and Sample Code
- PDF pp. 4–6: VI. Calibration and Synchronization, subsections A–D
- PDF pp. 6–7: VII. Evaluation Metrics, RPE and ATE
- PDF pp. 7–8: VIII. Conclusions, acknowledgements and references
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Separate supplemental material availability has not been fully verified.
- Identity: the observed title matches the catalog apart from capitalization. The title page verifies Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard and Daniel Cremers; the catalog author string renders Jürgen as Jrgen.
- Version: this eight-page PDF has no explicit revision identifier or publication-date imprint. Its immutable source hash identifies the reviewed artifact. Section II describes extensions of the 2011 workshop benchmark cited as reference [33]; that earlier work was not supplied or compared directly. No later release was substituted.
- The supplied text extraction does not reconstruct figure images; this gap was addressed by inspecting all eight PDF pages and all final crops. No appendix is present.
- Separate supplemental material availability has not been fully verified. Dataset archives, the benchmark website, linked code and external references were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author line and affiliation footnotes
The title and five authors are visible; Sturm and Cremers are affiliated with the Technical University of Munich, and Engelhard, Endres and Burgard with the University of Freiburg.
Go to primary source ↓e02PDF p. 1, abstract, Section I and Figure 1
Reports 39 sequences, 640 × 480 color/depth at 30 Hz, 100 Hz motion-capture ground truth, handheld/robot rigs and CC-BY 3.0 data licensing.
Go to primary source ↓e03PDF p. 2, Section II, earlier-benchmark paragraph; PDF p. 8, reference [33]
Describes extending the earlier benchmark with dynamic sequences, longer trajectories and robot recordings; reference [33] is the June 2011 workshop work Towards a benchmark for RGB-D SLAM evaluation.
Go to primary source ↓e04PDF p. 2, Section II, final paragraph; PDF p. 3, Section III, overview; PDF p. 4, Table I
Training sequences support offline evaluation; testing is described as website-only. Section III calls the 19 tabulated sequences training sequences.
Go to primary source ↓e05PDF p. 3, Section III(a–d); PDF p. 4, Section III continuation and Table I
Describes checkerboard calibration, isolated motions, moving objects, loop/no-loop tours, missing ground truth outside capture coverage and manually driven Pioneer sequences. Table I reports durations and average speeds.
Go to primary source ↓e06PDF p. 4, Sections IV–V
Documents the two Kinect units, Ubuntu 10.10/ROS Diamondback/OpenNI, factory-calibrated depth registration, eight Raptor-E cameras capable of up to 300 Hz, 500 Hz accelerometer logging and archive formats including depth scaling and pose quaternions.
Go to primary source ↓e07PDF p. 4, Table II; PDF p. 5, Section VI-B and Figure 4a
Lists color/infrared focal lengths and optical centers, depth correction ds = 1.035 and 1.031, and explains checkerboard calibration and that distributed depth is already corrected. Distortion calibration is mentioned but coefficients are not tabulated.
Go to primary source ↓e08PDF p. 5, Section VI-A; PDF p. 3, Figure 3
Cortex calibration uses a three-marker wand; checking an approximately two-metre rod gives a reported length standard deviation of 1.96 mm over the capture area.
Go to primary source ↓e09PDF p. 5, Section VI-C, both columns
Four Kinect markers and five board markers support extrinsic calibration. RGB/predicted corner residuals average 3.25/4.03 mm; marker-model residuals are 0.60/0.86 mm. The authors infer relative and absolute accuracy limits and warn against highly accurate map evaluation.
Go to primary source ↓e10PDF p. 6, Figure 5 and Section VI-D, first two paragraphs
Residual-versus-offset analysis finds motion capture approximately 20 ms earlier than color and says that delay is corrected. Depth arrives about 20 ms after color on average, but image timestamps remain unmodified; association accepts offset/tolerance preferences.
Go to primary source ↓e11PDF p. 6, Section VII, opening paragraphs in both columns
Evaluation uses estimated P and ground-truth Q camera poses in SE(3), referenced to the RGB optical frame, with association/interpolation needed in practice. The paper explicitly separates trajectory accuracy from map quality.
Go to primary source ↓e12PDF p. 6, Section VII-A, Eqs. (1)–(2); PDF p. 7, Section VII-A, Eq. (3) and adjacent paragraphs
Defines relative-motion discrepancy, m = n − Δ, translational RMSE, interval examples and sampled all-interval evaluation. Eq. (3) prints an upper limit n, creating a zero-pair endpoint under the preceding definition.
Go to primary source ↓e13PDF p. 7, Section VII-B, Eqs. (4)–(5)
Defines rigid least-squares Horn alignment S, absolute error F_i = Q_i^{-1}SP_i and translational RMSE over associated poses.
Go to primary source ↓e14PDF p. 6, Figure 6 and caption; PDF p. 7, Section VII-A, final paragraph
The fr1/desk plot compares RBM with GICP. The authors describe lower RBM drift and fewer outliers; no aggregate scores or exact plotted Δ are given.
Go to primary source ↓e15PDF p. 7, Figure 7, caption and Section VII-B discussion
Shows an aligned fr1/desk2 estimate and a log-log RPE/ATE comparison from RGB-D SLAM trajectories. The authors report strong correlation and often unchanged ordering, and assert a broader inequality without a proof here.
Go to primary source ↓e16PDF p. 5, Figure 4b and Section VI-B; PDF p. 6, Section VI-D, final paragraph
Depth standard deviation rises with distance. The prose approximates noise as 1 cm up to 2 m and 5 cm at 4 m; the latter is not an exact graph reading. Rolling shutter and automatic exposure can distort fast-motion imagery.
Go to primary source ↓e17PDF p. 4, Tables I–II; PDF pp. 6–7, Figures 6–7 and Section VIII
The tables describe sequence statistics and calibration, while algorithm results are illustrative plots. There is no proposed learned architecture, aggregate algorithm score table or controlled component ablation in the supplied paper.
Go to primary source ↓8.5 Primary sources
A Benchmark for the Evaluation of RGB-D SLAM Systems ↗
PDF · 7,200 extracted words
Source fingerprint
1f046af7c050b1edbef1122de6d322d01359f07b4278befbf8ccaf23c86d406d