REVIEW 4 major objections 3 minor 11 references
An Objectness Score for Accurate and Fast Detection during Navigation
T0 review · 4 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read A weighted objectness score keeps a mobile robot's detected object labels stable as it moves, by favoring views taken at sufficient distance.
desk verdict A plausible but unvalidated heuristic for arbitrating between conflicting class labels during navigation; the qualitative evaluation doesn't support the robustness claim. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central mechanism is the objectness score, a weighted sum $S_{obj} = \alpha S_{poc} + (1-\alpha)S_{depth}$ with $\alpha=0.4$; it treats $S_{depth}$ as a min-max normalized distance, capped at 0.8 m and 3 m, to favor views where the whole object shape is visible. It is paired with Algorithm 1, which uses k-nearest-neighbor search (k=3) and a 2D IoU threshold of 0.9 to decide that two detections are the same object, and with Eq. 1, the affine projection $p_{img} = K[R|t]P_{world}$, which maps stored 3D locations back into the current image. The score's role is to arbitrate between the stored class and a newly detected class at the same position.
What would settle it
Measure the class-label accuracy of Mask R-CNN on the same object across a continuous range of distances (for example, every 0.1 m from 0.3 to 3 m) and compare with the objectness score's ranking; if labels at the extremes are not worse than labels near the 0.8-3 m band, the score will systematically choose inferior labels.
Extended reading notes
Core claim
The central claim is that an objectness score $S_{obj}=\alpha S_{poc} + (1-\alpha)S_{depth}$ reliably decides which class label to keep for an object as a robot's viewpoint changes. $S_{poc}$ is the detector's class probability, and $S_{depth}$ is the distance between the camera and the object, normalized between 0.8 and 3 m. An object detected at a new pose is matched to an existing object by 2D intersection-over-union (threshold 0.9) among the k-nearest stored objects; if the new detection has a higher score, it replaces the stored one. The 3D position is saved in world coordinates and projected back to the image plane with an affine projection, so the comparison does not require a 3D detector or 3D IoU. The paper reports that this maintains object locations and classes in experiments at angles of -45, 0, and 45 degrees and distances from 0.3 to 3 m.
Load-bearing premise
The load-bearing premise is that a 2D object detector is most accurate when the object is at a 'sufficient distance'—close enough to see the whole shape but not so close that scale change confuses it; the paper states this property and illustrates it with one example but does not measure detector accuracy against distance.
Editorial extensions
If this is right
- A robot using this score can build a consistent semantic map from a 2D detector plus odometry and a calibrated camera, without 3D object detectors, which the paper reports are slower.
- The same stored 3D location can be re-projected into the image after the robot turns or moves, so object tracking survives large view changes.
- Choosing the higher score between overlapping detections means a wrong close-up label cannot overwrite a correct distant label, preventing the chair-to-traffic-light failure shown.
- Replacing stored information only when the new score is higher keeps the map stable over time and requires no retraining of the detector.
Reading between the lines
- Inference beyond the paper: the score's distance term is a proxy for detector accuracy; measuring actual class accuracy versus distance for Mask R-CNN would tell whether 0.8-3 m is the right band and could yield a data-driven $S_{depth}$.
- Inference beyond the paper: the same arbitration should apply to any 2D detector whose confidence is scale-dependent, including YOLOv3 and Faster R-CNN, but the paper only demonstrates it with Mask R-CNN.
- Inference beyond the paper: if a close-up view is the only one showing the true class, the distance weighting could freeze an earlier wrong label; this failure mode is testable by running the algorithm on sequences where the far view is ambiguous.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes an objectness score for maintaining the locations and class labels of objects detected by Mask R-CNN during mobile robot navigation. The score is a weighted combination of the detector's class probability and a distance-based term that is designed to favor observations made at a 'sufficient distance' from the object. Detected objects are stored in 3D world coordinates and projected back into 2D image coordinates via an affine projection; when a newly detected object overlaps a previously stored object, the class with the higher objectness score is retained. The authors evaluate the method in a Gazebo simulator by moving a turtlebot toward objects from different angles and distances, and report qualitative results showing that their algorithm maintains object classes better than raw Mask R-CNN outputs.
Significance. If the central assumption is validated, the method offers a lightweight alternative to 3D-IoU-based class selection in semantic navigation, with potential computational benefits. The paper identifies a real failure mode (scale change causing class misinterpretation) and proposes a simple, interpretable heuristic. However, the contribution is currently supported only by qualitative image sequences, and the key premise about distance-dependent detector accuracy is asserted rather than empirically established. The work does not include quantitative measurements, baselines, or machine-checked proofs, so its significance rests entirely on future validation.
major comments (4)
- [2.2, Eq. (2)] The objectness score is constructed so that S_depth increases monotonically with distance, but the paper never measures detector accuracy as a function of distance. This is load-bearing: when the robot approaches an object, the previously stored far-away observation has a higher S_depth, so Algorithm 1 tends to keep the farther label even if the closer view is actually more reliable. For example, with alpha=0.4, a far observation with S_depth=1.0 and class probability 0.7 yields S_obj=0.88, while a close observation with S_depth=0.0 and class probability 0.95 yields S_obj=0.38. The paper needs a quantitative study of class-conditional detector accuracy versus distance to support the monotonicity assumption, or the robustness claim is not justified.
- [3, Fig. 4] The experimental evidence consists of qualitative image sequences from a single simulator setup. There are no quantitative metrics such as label consistency, classification accuracy, or localization error; no comparison against baselines such as class-probability-only selection or 3D-IoU methods; and no repeated trials or error bars. The abstract's claim that the approach is 'efficient and robust, regardless of changing angles and distances' is therefore not supported. In addition, the claimed computational advantage over 3D IOU (120 ms) is never measured for the proposed method.
- [Algorithm 1, lines 4-11] The pseudocode is under-specified. SearchKNN is described as returning k=3 nearest neighbors, but line 4 assigns its result to the singular variable rho_i, and the while loop never updates rho_i. As written, the loop either does not iterate over the k neighbors or is infinite. This needs to be corrected and made precise (e.g., by iterating over the returned list with an index) because the algorithm is the core of the proposed method.
- [2.2, Eq. (2)] The text states that S_depth is a normalized distance with a value between 0 and 1, but the formula as written does not clamp d_i. In the experiments, the robot moves from 0.3 m to 3 m while min(d)=0.8 m, so for d_i=0.3 m, S_depth would be negative (0.3-0.8)/(3-0.8) = -0.227. The authors should either define clamping or explicitly state that negative values are allowed and explain the consequence for the objectness score.
minor comments (3)
- [3, first paragraph] The sentence 'The experimental results (Fig. 3) showed...' appears to reference the wrong figure; the qualitative results are shown in Fig. 4, so the citation should be corrected.
- [References] The citation to Rasouli and Tsotsos (2017) is a paper on color space selection for detectability, not a description of an indoor environment; please cite an appropriate environment or clarify why this reference supports the experimental setup.
- [Throughout] There are minor notation inconsistencies, such as P_oobj versus P_𝑜𝑏𝑗 for the list of detected objects, and the use of both Greek and Latin symbols for the same quantities; a pass to unify notation would improve readability.
Circularity Check
Distance-robustness is built into Eq. (2)–(3), so the reported challenge 'robust, regardless of changing angles and distances' is largely a consequence of the score's construction rather than an independent validation.
-
self definitional
[Abstract; Sec. 2.2, Eqs. (2)–(3); Alg. 1, Lines 8–9]
"Specifically, it is designed to increase when there is sufficient distance between a detected object and the camera. ... S_obj = αS_poc + (1−α)S_depth ... Sdepth = (d_i − min(d))/(max(d)−min(d)) ... if Check2DIOU(o_new, ρ_i) and S_o_new is larger than S_ρ_i then Replace ρ_i with o_new"
Eq. (2) defines S_depth as a min-max normalized distance that increases monotonically with distance over [0.8 m, 3 m]. Eq. (3) gives S_depth a weight of 1−α = 0.6 (with α = 0.4 set in Sec. 3). Alg. 1 replaces an existing object only when the new objectness score is larger. Therefore, when the robot approaches an object, the previously recorded farther observation necessarily has a higher S_depth and is retained whenever class probabilities are comparable. The paper's experimental conclusion that the class is 'maintained when the mobile robot moved toward the object' is thus a direct consequence of the definition of the score, not an independent test of the load-bearing premise that DLNs are more accurate at sufficient distance.
full rationale
The paper is not globally circular: the camera projection, KNN search, and 2D-IOU matching are standard components, and there is no load-bearing self-citation chain. The circularity is confined to the distance-robustness claim. The objectness score is explicitly designed so that S_depth increases with distance, and Algorithm 1 then retains the detection with the higher score. Consequently, demonstrating that the algorithm keeps the previously detected class while the robot approaches is a tautological execution of the construction. The underlying assumption that DLNs are more accurate at sufficient distance is never independently validated; the evaluation only shows that the algorithm behaves as designed. Because the central advertised result ('robust, regardless of changing angles and distances') depends on this unvalidated premise and the observed distance behavior is wired into Eqs. (2)–(3), the score is 6 rather than 0–2.
Assumptions & free parameters
free parameters (5)
- alpha =
0.4
- min(d) =
0.8 m
- max(d) =
3 m
- 2D IoU threshold =
0.9
- k =
3
assumptions (3)
- domain assumption Detector class probability is more reliable when the entire object is visible at a sufficient distance, so larger distance (up to max) implies better class estimates.
- domain assumption 2D IoU greater than 0.9 between a projected stored bounding box and a newly detected box reliably identifies the same physical object.
- domain assumption The affine projection p_img = K[R|t] P_world, with intrinsics from calibration and extrinsics from odometry, maps stored 3D positions into the current image accurately enough for IoU matching.
Cite this review
Pith. "Pith review of An Objectness Score for Accurate and Fast Detection during Navigation." pith.science (2026). https://pith.science/paper/O4XXZF5V
@misc{pith2026190905626,
author = {Pith},
title = {Pith review of: An Objectness Score for Accurate and Fast Detection during Navigation},
year = {2026},
howpublished = {\url{https://pith.science/paper/O4XXZF5V}},
note = {Machine review of arXiv:1909.05626}
}
read the original abstract
We propose a novel method utilizing an objectness score for maintaining the locations and classes of objects detected from Mask R-CNN during mobile robot navigation. The objectness score is defined to measure how well the detector identifies the locations and classes of objects during navigation. Specifically, it is designed to increase when there is sufficient distance between a detected object and the camera. During the navigation process, we transform the locations of objects in 3D world coordinates into 2D image coordinates through an affine projection and decide whether to retain the classes of detected objects using the objectness score. We conducted experiments to determine how well the locations and classes of detected objects are maintained at various angles and positions. Experimental results showed that our approach is efficient and robust, regardless of changing angles and distances.
Reference graph
Works this paper leans on
-
[1]
Probabilistic data association for semantic SLAM,
Bowman, S.L., Atanasov, N., Daniilidis, K., and Pappas, G.J. (2017), "Probabilistic data association for semantic SLAM,” IEEE Int. Conf. on Robotics and Automation, ICRA
work page 2017
-
[2]
He, K., Gkioxari, G., Dollar, P., and Girshick, R. (2017), "Mask R-CNN,” IEEE Int. Conf. on Computer Vision, ICCV, 2961-2969. Chair Chair Chair Umbrella Chair Umbrella Traffic light Traffic light Chair Chair Traffic light Not detected Chair Chair Chair Chair Chair Chair Chair Chair Chair
work page 2017
-
[3]
Accuracy and resolution of kinect depth data for indoor mapping applications,
Khoshelham, K. and Elberink, S.O. (2012), "Accuracy and resolution of kinect depth data for indoor mapping applications,” Sensors, Vol. 12(2), 1437-1454. Kim J.Y. and Yoon S.E. (2018), "Regional Attention Based Deep Feature for Image Retrieval,” British Machine Vision Conf., BMVC
work page 2012
-
[4]
Super Rays and Culling Region for Real-Time Updates on Grid-based Occupancy Maps,
Kwon, Y.S., Kim D.H., An I.K., and Yoon S.E. (2019), "Super Rays and Culling Region for Real-Time Updates on Grid-based Occupancy Maps,” IEEE Transaction on
work page 2019
-
[5]
Globally -optimal greedy algorithms for tracking a variable number,
Pirsiavash, H., Ramanan, D., and Fowlkes, C.C. (2011), "Globally -optimal greedy algorithms for tracking a variable number,” IEEE Conf. on Computer Vision and Pattern Recognition, CVPR
work page 2011
-
[6]
The effect of color space selection on detectability and discriminability of colored objects,
Rasouli, A. and Tsotsos, J.K. (2017), "The effect of color space selection on detectability and discriminability of colored objects,” arXiv:1702
work page 2017
-
[7]
YOLOv3: An incremental improvement,
Redmon, J. and Farhadi, A. (2018), "YOLOv3: An incremental improvement,” arXiv:1804
work page 2018
-
[8]
Faster R-CNN: towards real-time object Detection with region proposal networks,
Ren, S., He, K., Girshick, R., and Sun, J. (2015), "Faster R-CNN: towards real-time object Detection with region proposal networks,” Neural Information Processing Systems , NIPS
work page 2015
Show all 11 references
-
[9]
Deep sliding shapes for amodal 3D object detection in RGB-D images,
Song, S. and Xiao, J. (2016), “Deep sliding shapes for amodal 3D object detection in RGB-D images,” IEEE Conf. on Computer Vision and Pattern Recognition, CVPR
2016
-
[10]
Frustum pointnets for 3D object detection from RGB -D data,
Qi, C.R., Liu, W., Wu, C., Su, H., and Guibas, L.J. (2018), “Frustum pointnets for 3D object detection from RGB -D data,” IEEE Conf. on Compu ter Vision and Pattern
2018
-
[11]
A flexible new technique for camera calibration,
Zhang, Z. (2000), "A flexible new technique for camera calibration,” in IEEE Transaction on Pattern Analysis and Machine Intelligence, TPAMI, Vol. 22, 1330-1334
2000
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.