Pith. sign in

REVIEW 3 major objections 5 minor 47 references

A robot can learn to estimate another robot's full pose (position, distance, bearing) from monocular RGB images using only binary LED on/off labels as supervision — no pose labels, no CAD model, no motion capture.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 17:52 UTC pith:GR32TROD

load-bearing objection Clever and sound in-lab result—pose from LED-state labels alone works—but the bearing mechanism rests on an unvalidated cosine visibility model, and distance/OOD claims are weaker than advertised. the 3 major comments →

arxiv 2509.10405 v1 pith:GR32TROD submitted 2025-09-12 cs.RO

Self-supervised Learning Of Visual Pose Estimation Without Pose Labels By Classifying LED States

Nicholas Carlotti , Mirko Nava , Alessandro Giusti This is my paper
classification cs.RO
keywords self-supervised learningvisual pose estimationpretext taskLED state classificationrelative robot posemonocular RGBmulti-robot localization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper claims that full relative pose — where the peer robot is in the image, how far away it is, and which way it faces — can be learned from scratch without any pose labels. The only supervision is the binary on/off state of four LEDs on the target robot, known at each training frame. The network trains on the pretext task of classifying each LED's state; to succeed it must implicitly locate the robot, determine which side is visible, and gauge its apparent size. At inference the LEDs are not needed and can be anything. The authors report the self-supervised model closely matches a fully supervised upper bound and outperforms CAD-based baselines on distance and heading.

Core claim

The paper's core discovery is that the multi-LED state classification pretext task carries enough information to teach a convolutional network the geometry of the robot. The model outputs per-cell predictions: LED-state confidence, robot-presence probability, and bearing angle. The training loss multiplies the LED classification error by a softmax-normalized presence map (so cells that see the robot dominate) and by a cosine visibility function that says which LEDs should be visible for any predicted bearing. To minimize the loss the network must locate the robot, infer the viewing direction, and understand its body structure. At test time, pose is read off the same maps. The distance comes

What carries the argument

The mechanism is the spatially-weighted, visibility-modulated classification loss. A fully convolutional network (a network made only of convolutions and pooling, so each output cell corresponds to a local receptive field) produces, for each cell, a predicted LED state, a presence score, and a bearing. The loss combines binary cross-entropy on each LED with (a) a softmax-normalized presence map that forces the model to attend only to cells whose receptive field contains the robot, and (b) a fixed cosine visibility model Λ_k = cos(ψ + 2π(k−1)/K) that downweights LEDs the robot's own body would occlude. The same forward pass is run at three image scales; the softmax over scales doubles as a di

Load-bearing premise

The load-bearing assumption is that the cosine visibility function accurately describes which LEDs are visible from any camera angle; the paper states this is approximate and not calibrated, and if real LED visibility deviates, bearing learning fails and the whole pose estimate degrades.

What would settle it

Collect a dataset where a robot with four LEDs is photographed from many known bearings, with one LED lit at a time. If the measured visibility of each LED as a function of bearing deviates sharply from a cosine (e.g., an LED stays fully visible well beyond the angle where the cosine predicts it should fade), then training the proposed loss on this data will produce systematically biased bearing estimates, and the reported accuracy will not be reproducible.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A pair of robots can collect training data autonomously: they move randomly, toggle LEDs, and broadcast their states; no external infrastructure or human labeling is required.
  • Pose estimation works at deployment even when all LEDs are off or arbitrarily set, so the method is not a LED-detection shortcut.
  • The trained model transfers to new environments without fine-tuning, and fine-tuning on a small amount of new-domain data outperforms training from scratch on that same data.
  • The method supports multi-robot pose estimation by interpreting each peak in the presence map as a separate robot, despite being trained with at most one visible robot.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The supervisory channel need not be LEDs: any appearance change that is observable and known during data collection — an articulated arm position, a screen pattern, a projected light — could in principle provide the same self-supervision. This is a natural generalization the paper hints at but does not develop.
  • The fixed cosine visibility function is the most fragile hand-designed element. A learned or calibrated visibility model would likely improve robustness to non-ideal LED emission patterns and partial occlusions, and could be tested as a drop-in replacement.
  • The distance estimate is inherently quantized by the number and spacing of image scales; the step-function artifact visible in the authors' plots suggests that a continuous scale regression or a feature-pyramid architecture would give smoother, more accurate distance predictions.
  • The method could transfer to cooperative marker-based localization in other domains — e.g., drones, vehicles, or surgical tools — as long as the marker states are known and one calibration image is available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a method to train a monocular RGB relative pose estimator for a ground robot without pose labels or a CAD model. The training signal is a pretext task: given an image from an observing robot, predict the binary on/off state of K LEDs on the target robot, for which states are broadcast and known. A fully convolutional network outputs spatial maps for LED states, robot presence, and relative bearing. The loss is a BCE term per LED, spatially weighted by a softmax-normalized presence map, and bearing-weighted by a cosine visibility model over the predicted bearing. Multi-scale input passes allow the model to select the scale at which the robot best fits the receptive field; that selection is converted to a metric distance via a single calibration factor. Experiments on a DJI RoboMaster S1 with 131K lab images (23% containing a robot), motion-capture ground truth, and three replicas show localization and bearing errors close to a supervised upperbound (17 px / 17° vs 18 px / 14°), while distance error is larger (24% vs 11%). The paper also reports OOD generalization, fine-tuning, and qualitative multi-robot results.

Significance. The contribution is original and practically relevant. It removes the need for pose labels or CAD models, replacing them with autonomously generated binary LED labels, and the architecture is simple and efficient (153 Hz). Strengths include a controlled experimental protocol with motion capture, a supervised upperbound, multiple replicas, and explicit limitations. The core idea is falsifiable and the paper provides enough detail to reproduce. The main risk is the uncalibrated cosine visibility model that carries the bearing-learning signal; if it deviates from the real LED visibility pattern, the learned bearing can be biased. The evidence that it works on this particular robot is good, but the assumption's robustness is not demonstrated. Distance resolution is also coarse. With additional calibration or sensitivity analysis, the method would be a solid contribution.

major comments (3)
  1. [Section 3, Eq. (1), Figure 3] The only supervision for the relative bearing ψ comes from the weighting of LED-state BCE terms by Λ_k = cos(ψ + 2π(k−1)/K). The authors state that this visibility function is 'not precise nor the result of calibration.' If the actual visibility pattern of the four LEDs deviates from this cosine model—e.g., because of body occlusion geometry or non-equidistant LED mounting—the loss can be minimized by a ψ that does not correspond to the true bearing. The reported Eψ=17° on this robot is encouraging, but it does not establish robustness. Please add a calibration of Λ for the actual robot, or a sensitivity analysis that perturbs Λ (angular offsets, widths, number of LEDs) and reports Eψ. Without this, the central bearing-learning mechanism rests on an unmeasured assumption.
  2. [Section 5, Table 1, Figure 6] Distance estimation is a core output of the pose. The reported Ed=24% versus 11% for the supervised upperbound is substantially worse, and the text attributes this to the three-scale discretization. The claim that 'the issue can be mitigated by using more scales' is only mentioned as a preliminary experiment without reporting the result. Please provide a quantitative evaluation with a finer scale set (or an analysis of the distance error as a function of the number of scales) to support the claim that the method can be competitive on all pose components. At minimum, qualify the 'competitive' claim to localization and bearing.
  3. [Section 5, 'The Model is Capable of Multi-robot Pose Estimation' and Appendix C] The multi-robot capability is listed as a contribution and a main result, but the only support is the qualitative examples in Figure 8. The inference modification in Appendix C changes the softmax behavior and is non-trivial; it deserves a quantitative evaluation (e.g., detection accuracy, pose error on images with two or more robots). If this is intended as a demonstration, the text should say so and place it in context rather than as a headline result.
minor comments (5)
  1. [Section 3 / Eq. (4) / Appendix A] The notation for scale factors is inconsistent. Appendix A lists input rescaling factors as (1, 1/2, 1/4), but Eq. (4) uses f_s and the text says 'average of the inverse of scale factors.' Clarify whether f_s are the inverse scales (1, 2, 4) and define the variable consistently throughout.
  2. [Table 1] The table reports 'three replicas per row' but only single values are shown for each metric; the 95% CI is only visible in the point plot. State how the three replicas were aggregated and report uncertainties for the numeric columns.
  3. [Section 5 / Figure 7] The fine-tuning comparison has no error bars or number of replicas; adding them would strengthen the conclusion.
  4. [Related Work] Refs. [38] and [39] are from the same group and are closely related. The text distinguishes them by the lack of fine-tuning, but a more explicit comparison—architectural changes and experimental differences—would help the reader understand the novelty.
  5. [Section 4 / Appendix D] The robot detection AUC (97.2% with LED confidence) is an interesting byproduct but is only in the appendix. Consider mentioning it in the main text as a further benefit, or remove it if it distracts from the pose estimation focus.

Circularity Check

0 steps flagged

No significant circularity: pose predictions are learned from an independent LED-state supervision signal; the cosine visibility model and one-point distance calibration are explicit assumptions, not fitted pose labels.

full rationale

The paper's derivation chain is self-contained with respect to the pose predictions. Training supervision is the binary state of four LEDs, broadcast by the target robot; pose is never used as a label for the proposed model. Localization emerges from a spatial weighting of the LED-classification loss (softmax-normalized P map), bearing emerges from an assumed cosine visibility function Lambda_k, and distance is inferred from a multi-scale selection of the best-fitting receptive field, calibrated by a single known-distance image (Eq. 4). None of these outputs are defined as functions of the pose labels; the only external numeric input is the single scalar d_c, which fixes metric scale and does not determine per-image distance or bearing. The cosine visibility function is explicitly acknowledged as an approximation ('This visibility function reasonably approximates each LED's visibility from different viewing directions, though it is not precise nor the result of calibration'), and the paper validates bearing against motion-capture ground truth (E_psi = 17 deg), so the bearing claim is empirically supported rather than true by construction. The self-citations [38,39] are contextual and not load-bearing: the present method explicitly contrasts itself with those fine-tuning approaches. Therefore no step reduces to its own input.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

The method introduces no new physical entities. Its load-bearing assumptions are the cosine LED-visibility model, the known LED directions, and the single-image metric calibration. These are explicit in the paper, and the authors flag the visibility and scale discretization as approximate.

free parameters (2)
  • d_c (monocular distance calibration factor) = single calibration image at known distance (Appendix B)
    Converts apparent-size estimate, a weighted inverse-scale combination, into metric distance; resolves monocular depth scale ambiguity. Disclosed as calibration.
  • image scale set {1, 1/2, 1/4} = chosen by hand
    Distance is estimated by selecting among these three discrete scales; coarse discretization limits distance accuracy (Ed=24%). Authors admit more scales would help.
axioms (6)
  • domain assumption LED state changes observably affect the robot's appearance in the camera image at the known approximate viewing directions
    Section 1: 'The only assumption is that the LED state affects the robot's appearance in a way that is observable from an image...'
  • domain assumption Each LED's approximate viewing direction relative to robot heading is known
    Training-time assumption (i) in the abstract and Section 1.
  • ad hoc to paper LED visibility from a viewpoint follows Lambda_k = cos(psi + 2*pi*(k-1)/K) for K equidistant LEDs
    Section 3 and Figure 3: authors call it an approximation, 'not precise nor the result of calibration'.
  • domain assumption Apparent robot size in the image is inversely proportional to distance, and the chosen image scale fixes the metric scale when multiplied by d_c
    Section 3, Robot Distance Estimation and Appendix B; standard perspective projection.
  • domain assumption At most one robot appears in the field of view during training
    Section 7 Limitations: 'The training approach assumes to have at most one robot inside the FOV.'
  • standard math Standard CNN backpropagation and softmax weighting produce a non-degenerate P map
    Section 3; softmax prevents the model from trivially zeroing the localization loss.

pith-pipeline@v1.3.0-alltime-deepseek · 12611 in / 14792 out tokens · 151837 ms · 2026-08-04T17:52:31.565464+00:00 · methodology

0 comments
read the original abstract

We introduce a model for monocular RGB relative pose estimation of a ground robot that trains from scratch without pose labels nor prior knowledge about the robot's shape or appearance. At training time, we assume: (i) a robot fitted with multiple LEDs, whose states are independent and known at each frame; (ii) knowledge of the approximate viewing direction of each LED; and (iii) availability of a calibration image with a known target distance, to address the ambiguity of monocular depth estimation. Training data is collected by a pair of robots moving randomly without needing external infrastructure or human supervision. Our model trains on the task of predicting from an image the state of each LED on the robot. In doing so, it learns to predict the position of the robot in the image, its distance, and its relative bearing. At inference time, the state of the LEDs is unknown, can be arbitrary, and does not affect the pose estimation performance. Quantitative experiments indicate that our approach: is competitive with SoA approaches that require supervision from pose labels or a CAD model of the robot; generalizes to different domains; and handles multi-robot pose estimation.

Figures

Figures reproduced from arXiv: 2509.10405 by Alessandro Giusti, Mirko Nava, Nicholas Carlotti.

Figure 1
Figure 1. Figure 1: By solving the multi-LED state classification task ( [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the approach: (a) given an input image, our approach predicts the robot’s [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visibility function for a robot with four LEDs at the cardinal directions. ψˆ is a cell of the predicted bearing map Ψˆ , and λˆk is the visibility weight for each of the K LEDs. Robot Relative Bearing Estimation. Whenever the robot is visible in the image, some of its LEDs are occluded by its own body. Since these LEDs are not visible, the model is unable to predict their state, contributing to high value… view at source ↗
Figure 4
Figure 4. Figure 4: Model’s output maps (a single scale was selected for visualization purposes). The [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Random training samples from the datasets: laboratory (top row), gym (1-3 on bottom), [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Our self-supervised model pre￾dictions vs ground truth on Qν lab. Our ap￾proach discretizes distances into three bins, resulting in a coarse step function. Our model’s performance largely surpasses the Mean Predictor and closely follows the Upperbound. The performance gap with the latter is in the distance es￾timation; this is explained by the fully-supervised ap￾proach regressing the robot’s distance as a… view at source ↗
Figure 8
Figure 8. Figure 8: Predicted robot poses (x-axis in red, y-axis in green, z-axis in blue) by models trained with our approach: a model trained on Tlab applied to Qν lab (1-4); a model trained on Tood and applied to the testing set in classroom (5, 6), gym (7), and break room (8); a model trained on Tlab applied to images with multiple robots (9, 10). Large errors occur when the images are blurred or robots are far from the c… view at source ↗
Figure 7
Figure 7. Figure 7: Metrics of a model trained from scratch on Tlab (in green) and a model pre￾trained on Tood and fine-tuned on Tlab (in yel￾low). The performance of Mean Predictor (in gray) and Upperbound (in magenta) are reported as dashed lines for comparison. We further fine-tune the model on the target domain, represented by the Fine-tuning model, and compare with another trained from scratch on the same data, called Fr… view at source ↗
Figure 9
Figure 9. Figure 9: Receiver Operator Characteristic (ROC) curves for the robot detection methods presented [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 3 linked inside Pith

  1. [1]

    Dorigo, G

    M. Dorigo, G. Theraulaz, and V . Trianni. Swarm robotics: Past, present, and future.IEEE Point of View, 109(7):1152–1165, 2021

    2021

  2. [2]

    G. Wang, F. Manhardt, F. Tombari, and X. Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16611–16621, 2021

    2021

  3. [3]

    X. Liu, R. Zhang, C. Zhang, G. Wang, J. Tang, Z. Li, and X. Ji. Gdrnpp: A geometry-guided and fully learning-based object pose estimator.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  4. [4]

    Jing and Y

    L. Jing and Y . Tian. Self-supervised visual feature learning with deep neural networks: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020

    2020

  5. [5]

    Salvato, G

    E. Salvato, G. Fenu, E. Medvet, and F. A. Pellegrino. Crossing the reality gap: A survey on sim-to-real transferability of robot controllers in reinforcement learning.IEEE Access, 9: 153171–153187, 2021

    2021

  6. [6]

    V . N. Nguyen, T. Groueix, G. Ponimatkin, V . Lepetit, and T. Hodan. Cnos: A strong baseline for cad-based novel object segmentation. InIEEE/CVF International Conference on Computer Vision, pages 2134–2140, 2023

    2023

  7. [7]

    Ausserlechner, D

    P. Ausserlechner, D. Haberger, S. Thalhammer, J.-B. Weibel, and M. Vincze. Zs6d: Zero- shot 6d object pose estimation using vision transformers.IEEE International Conference on Robotics and Automation, 2023

    2023

  8. [8]

    S. Moon, H. Son, D. Hur, and S. Kim. Genflow: Generalizable recurrent flow for 6d pose refinement of novel objects. InIEEE/CVF Conference on Computer Vision and Pattern Recog- nition, pages 10039–10049, 2024

    2024

  9. [9]

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371–10381, 2024

    2024

  10. [10]

    L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao. Depth anything v2. Advances in Neural Information Processing Systems, 37:21875–21911, 2024

    2024

  11. [11]

    Xiang, T

    Y . Xiang, T. Schmidt, V . Narayanan, and D. Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes.Robotics: Science and Systems, 2018

    2018

  12. [12]

    H. X. Pham, A. Sarabakha, M. Odnoshyvkin, and E. Kayacan. Pencilnet: Zero-shot sim-to- real transfer learning for robust gate perception in autonomous drone racing.IEEE Robotics and Automation Letters, 7(4):11847–11854, 2022

    2022

  13. [13]

    S. Li, C. De Wagter, and G. C. De Croon. Self-supervised monocular multi-robot relative localization with efficient deep neural networks. InIEEE International Conference on Robotics and Automation, pages 9689–9695, 2022

    2022

  14. [14]

    X. Deng, Y . Xiang, A. Mousavian, C. Eppner, T. Bretl, and D. Fox. Self-supervised 6d object pose estimation for robot manipulation. InIEEE International Conference on Robotics and Automation, pages 3665–3671, 2020

    2020

  15. [15]

    T. E. Lee, J. Tremblay, T. To, J. Cheng, T. Mosier, O. Kroemer, D. Fox, and S. Birchfield. Camera-to-robot pose estimation from a single image. InIEEE International Conference on Robotics and Automation, pages 9426–9432, 2020

    2020

  16. [16]

    Y . Su, M. Saleh, T. Fetzer, J. Rambach, N. Navab, B. Busam, D. Stricker, and F. Tombari. Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6738–6748, 2022. 10

    2022

  17. [17]

    Labb´e, L

    Y . Labb´e, L. Manuelli, A. Mousavian, S. Tyree, S. Birchfield, J. Tremblay, J. Carpentier, M. Aubry, D. Fox, and J. Sivic. Megapose: 6d pose estimation of novel objects via render & compare. InPMLR Conference on Robot Learning, pages 715–725, 2023

    2023

  18. [18]

    Kirillov, E

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, et al. Segment anything. InIEEE/CVF International Conference on Computer Vision, pages 4015–4026, 2023

    2023

  19. [19]

    Oquab, T

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haz- iza, F. Massa, A. El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    Pith/arXiv arXiv 2023

  20. [20]

    J. Lin, L. Liu, D. Lu, and K. Jia. Sam-6d: Segment anything model meets zero-shot 6d object pose estimation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27906–27916, 2024

    2024

  21. [21]

    V . N. Nguyen, T. Groueix, M. Salzmann, and V . Lepetit. Gigapose: Fast and robust novel object pose estimation via one correspondence. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9903–9913, 2024

    2024

  22. [22]

    E. P. ¨Ornek, Y . Labb´e, B. Tekin, L. Ma, C. Keskin, C. Forster, and T. Hodan. Foundpose: Unseen object pose estimation with foundation features. InEuropean Conference on Computer Vision, pages 163–182, 2024

    2024

  23. [23]

    Mildenhall, P

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. InEuropean Conference on Computer Vision, 2020

    2020

  24. [24]

    Kerbl, G

    B. Kerbl, G. Kopanas, T. Leimk ¨uhler, and G. Drettakis. 3d gaussian splatting for real-time radiance field rendering.ACM Transactions on Graphics, 42(4), 2023

    2023

  25. [25]

    Yen-Chen, P

    L. Yen-Chen, P. Florence, J. T. Barron, A. Rodriguez, P. Isola, and T.-Y . Lin. iNeRF: Inverting neural radiance fields for pose estimation. InIEEE/RSJ International Conference on Intelligent Robots and Systems, 2021

    2021

  26. [26]

    Bortolon, T

    M. Bortolon, T. Tsesmelis, S. James, F. Poiesi, and A. Del Bue. Iffnerf: Initialisation free and fast 6dof pose estimation from a single image and a nerf model.arXiv preprint arXiv:2403.12682, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Matteo, T

    B. Matteo, T. Tsesmelis, S. James, F. Poiesi, and A. Del Bue. 6dgs: 6d pose estimation from a single image and a 3d gaussian splatting model. InEuropean Conference on Computer Vision, pages 420–436, 2024

    2024

  28. [28]

    Zhang, J

    D. Zhang, J. Han, G. Cheng, and M.-H. Yang. Weakly supervised object localization and detection: A survey.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9): 5866–5885, 2021

    2021

  29. [29]

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba. Learning deep features for discriminative localization. Incvpr, pages 2921–2929, 2016

    2016

  30. [30]

    H. G. Ramaswamy et al. Ablation-cam: Visual explanations for deep convolutional network via gradient-free localization. InIEEE/CVF winter conference on applications of computer vision, pages 983–991, 2020

    2020

  31. [31]

    Jiang, C.-B

    P.-T. Jiang, C.-B. Zhang, Q. Hou, M.-M. Cheng, and Y . Wei. Layercam: Exploring hierarchical class activation maps for localization.IEEE Transactions on Image Processing, 30:5875–5888, 2021

    2021

  32. [32]

    W. Lu, X. Jia, W. Xie, L. Shen, Y . Zhou, and J. Duan. Geometry constrained weakly supervised object localization. InEuropean Conference on Computer Vision, pages 481–496, 2020. 11

    2020

  33. [33]

    J. Xie, J. Xiang, J. Chen, X. Hou, X. Zhao, and L. Shen. C2am: Contrastive learning of class- agnostic activation map for weakly supervised object localization and semantic segmentation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989–998, 2022

    2022

  34. [34]

    E. Jang, C. Devin, V . Vanhoucke, and S. Levine. Grasp2vec: Learning object representations from self-supervised grasping. InPMLR Conference on Robot Learning, pages 99–112, 2018

    2018

  35. [35]

    Radosavovic, T

    I. Radosavovic, T. Xiao, S. James, P. Abbeel, J. Malik, and T. Darrell. Real-world robot learning with masked visual pre-training. InPMLR Conference on Robot Learning, pages 416–426, 2023

    2023

  36. [36]

    Antsfeld and B

    L. Antsfeld and B. Chidlovskii. Self-supervised pretraining and finetuning for monocular depth and visual odometry. InIEEE International Conference on Robotics and Automation, pages 14669–14676, 2024

    2024

  37. [37]

    S. Qian, K. Mo, V . Blukis, D. F. Fouhey, D. Fox, and A. Goyal. 3d-mvp: 3d multiview pretraining for robotic manipulation.arXiv preprint arXiv:2406.18158, 2024

    Pith/arXiv arXiv 2024

  38. [38]

    M. Nava, N. Carlotti, L. Crupi, D. Palossi, and A. Giusti. Self-supervised learning of visual robot localization using led state prediction as a pretext task.IEEE Robotics and Automation Letters, 9(4):3363–3370, 2024

    2024

  39. [39]

    Carlotti, M

    N. Carlotti, M. Nava, and A. Giusti. Learning to estimate the pose of a peer robot in a camera image by predicting the states of its leds. InIEEE/RSJ International Conference on Intelligent Robots and Systems, pages 2763–2769, 2024

    2024

  40. [40]

    J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3431–3440, 2015

    2015

  41. [41]

    Guo, T.-X

    M.-H. Guo, T.-X. Xu, J.-J. Liu, Z.-N. Liu, P.-T. Jiang, T.-J. Mu, S.-H. Zhang, R. R. Martin, M.-M. Cheng, and S.-M. Hu. Attention mechanisms in computer vision: A survey.Springer Computational visual media, 8(3):331–368, 2022

    2022

  42. [42]

    K. V . Mardia and P. E. Jupp.Directional statistics. John Wiley & Sons, 2009

    2009

  43. [43]

    Y . Di, F. Manhardt, G. Wang, X. Ji, N. Navab, and F. Tombari. So-pose: Exploiting self- occlusion for direct 6d pose estimation. InIEEE/CVF International Conference on Computer Vision, pages 12396–12405, 2021

    2021

  44. [44]

    Ronneberger, P

    O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical im- age segmentation. InSpringer Medical image computing and computer-assisted intervention, pages 234–241, 2015

    2015

  45. [45]

    M. Nava, L. M. Gambardella, and A. Giusti. State-consistency loss for learning spatial per- ception tasks from partial labels.IEEE Robotics and Automation Letters, 6(2):1112–1119, 2021

    2021

  46. [46]

    D. P. Kingma and J. Ba. Adam: a method for stochastic optimization. InInternational Con- ference on Learning Representations, 2015

    2015

  47. [47]

    Loshchilov and F

    I. Loshchilov and F. Hutter. Sgdr: Stochastic gradient descent with warm restarts. InInterna- tional Conference on Learning Representations, 2017. 12 Appendix A Neural Network Training Throughout this document, we adopt the same FCN architecture, receiving as input an image with an original resolution of640×360pixels and producing maps of80×45cells, with ...

    2017