acceptodds
Withdrawn from ICLR 2027

High Performance Space Debris Tracking in Complex Skylight Backgrounds with a Large-Scale Dataset

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

With the rapid development of space exploration, space debris has attracted more attention due to its potential extreme threat, leading to the need for real-time and accurate debris tracking. However, existing methods are mainly based on traditional signal processing, which cannot effectively process the complex background and dense space debris. In this paper, we propose a deep learning-based Space Debris Tracking Network(SDT-Net) to achieve highly accurate debris tracking. SDT-Net effectively represents the feature of debris, enhancing the efficiency and stability of end-to-end model learning. To train and evaluate this model effectively, we also produce a large-scale dataset Space Debris Tracking Dataset(SDTD) by a novel observation-based data simulation scheme. SDTD contains 18,040 video sequences with a total of 63,247 frames and covers 27,350 synthetic space debris. Extensive experiments validate the effectiveness of our model and the challenging of our dataset. Furthermore, we annotate a real debris observation test set to evaluate the generalization of the model from simulated data to real observation scenarios. SDT-Net achieves a MOTA score of 76% on this set, which demonstrates its transferability to real-world scenarios. Our dataset and code will be released soon.

Primary area:
datasets and benchmarks
Keywords:
Benchmark, Tracking, AI for Science

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Q6hzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

To address the challenge of debris tracking in complex scenarios within space environments, the authors propose a sample-rich debris tracking dataset, termed SDTD. Building upon this dataset, they construct a space debris tracking model, SDT-Net, which incorporates a DAE module to highlight debris features. Subsequently, a "debris offset" module is employed for cross-frame data association to generate a tracking trajectory for each debris object. In addition, the authors propose a DKA module, which exploits the physical spatial continuity of debris motion across frames to achieve accurate and efficient association. Relatively comprehensive comparative and ablation experiments demonstrate the value of the proposed dataset and the effectiveness of the proposed modules.

Critical strengths

(1) The authors appear to have proposed a relatively data-rich space debris tracking dataset, SDTD, whose release will contribute to a certain extent to the development of the community.
(2) The authors conduct relatively comprehensive comparative and ablation experiments, which can, to a certain extent, demonstrate the effectiveness of the proposed model on this task.

Critical weaknesses

(1) The authors employ simulation tools to generate space debris. How do the authors ensure that the motion patterns and illumination of the simulated debris data remain consistent with those in real space scenarios? Should the authors not compare and analyze the simulated data against real data across multiple dimensions? For example, the authors assume the trajectory motion of debris to be linear motion. What is the theoretical basis for this assumption? The authors should provide a rigorous derivation from orbital kinematics to justify it.
(2) The methods compared in Table 2 are all from three years ago, whereas the authors also note in the paper that several papers on space debris tracking have appeared in the past two years. The authors should incorporate comparison results with these more recent methods. In addition, the fonts in figures such as Figure 5 and Figure 7 are too small, and there is overlapping and blurring, making them difficult for readers to read.
(3) In the SDT-Net network proposed by the authors, the allegedly innovative DAE and DKA modules are both rather common feature fusion and interaction operations in object tracking. Moreover, the overall network framework design adopts a very classic tracking network model, resembling the idea of an RNN model, so the novelty of the entire network is very limited.

Questions for authors

Please refer to the content of the Weakness section and provide detailed responses to each of the issues raised.

Reviewer_GGF3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces SDTD, a dataset of synthetic debris trajectories on real astronomical backgrounds, and SDT-Net for tracking line-shaped targets. The model combines endpoint detection and pairing, segmentation-based feature enhancement, and geometric and motion-based association. Experiments include sparse and dense synthetic sequences, component ablations, and a real-observation test set of 50 videos.

Critical strengths

1: The dataset addresses a useful application where annotated tracking data are limited. Combining real astronomical backgrounds with synthetic debris provides a practical way to generate training data, and the real-observation test set adds useful validation. 2: The reported improvements are supported by several comparisons and ablations. On the real test set, SDT-Net achieves 76.0 MOTA compared with 71.6 for OC-SORT. The appendix also includes newer trackers, cross-background tests, and longer-sequence training experiments.

Critical weaknesses

1: The motion simulation needs better justification. Appearance similarity is examined in detail, but the choice of constant-speed straight trajectories and the simulation parameter ranges are not sufficiently justified. Since DKA also uses a constant-velocity assumption, the method may benefit from closely matching the generator. The real test results are encouraging, but limited information about data collection, annotation, and test independence makes the transfer claim harder to assess.

2: The technical contribution and comparison protocol need clarification. The method combines fairly standard detection, feature enhancement, and association components, without clearly explaining the new insight beyond their adaptation to line-shaped targets. Baseline detector configurations and task-specific adaptations are also insufficiently described. In particular, the paper does not clearly explain how endpoint predictions and bounding-box predictions are evaluated consistently, which matters for interpreting the gains.

Questions for authors

1: How were the motion assumptions and simulation parameter ranges chosen? What evidence supports their agreement with real temporal observations, beyond appearance similarity? 2: What technical insight does SDT-Net add beyond combining endpoint pairing, segmentation-based enhancement, and geometric association? 3: Which detectors, training configurations, and task-specific adaptations were used for each baseline? The statement that official settings were followed does not fully explain these choices. 4: Where did the real videos come from, how were they annotated, and were they kept separate from training, tuning, and model selection? 5: How are line endpoints and bounding boxes converted into a common evaluation representation? Please specify the matching criteria and thresholds.

Reviewer_krmAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SDT-Net, a network for space debris tracking, and constructs a synthetic tracking dataset, SDTD, based on ZTF astronomical images. The method represents space debris in long-exposure images as line segments and performs line segment detection and endpoint pairing using endpoint heatmaps and embedding features. It also employs offset prediction to establish cross-frame associations. In addition, the network incorporates a DAE segmentation enhancement module and a DKA geometry- and motion-based association module to improve detection and tracking performance in complex backgrounds. Experimental results demonstrate that SDT-Net outperforms the comparison methods on both the sparse and dense test sets of SDTD. On a real-observation subset comprising 50 video sequences and 3,190 frames, the method achieves a multiple object tracking accuracy (MOTA) of 76.0%. The main contributions of this work are the construction of a synthetic dataset tailored to space debris tracking and the development of a corresponding tracking method that integrates line segment representations, geometric information, and motion features.

Critical strengths

Space debris in long-exposure images typically appears as streaks. Endpoint prediction is therefore better suited to the morphological characteristics of these targets than conventional bounding-box detection. If the reliability of sequence identities, data splits, and real-world observation results can be thoroughly validated, an astronomical background dataset with consecutive-frame annotations would be valuable for facilitating further research in this area.

Critical weaknesses

The main contribution relies on synthetic data, but the correspondence between the synthetic data generation process and real-world observations has not been established. The simulation described in the paper places line sources on ZTF backgrounds and incorporates several steps, including constant-velocity motion, rectangular line segments, random brightness perturbations, and PSF convolution. However, the paper does not adequately quantify how the distributions of key parameters are estimated from real debris observations, how the simulated PSF, exposure settings, and noise correspond to the actual imaging process, or how each simulation factor affects the results. Visual similarity between synthetic and real images, or similarity in their intensity distributions, is insufficient to demonstrate that the training distribution adequately covers the variations in the motion and imaging characteristics of real space debris. If the synthetic dataset is not representative of real observations, the substantial improvements reported on synthetic test sets cannot substantiate the claim that the proposed method is effective for real-world space debris tracking. The evaluation on real-world data is insufficient to support strong claims of generalization. SDTD-Real contains only 50 video sequences and 3,190 frames, yet the paper does not adequately describe the video sources, observing conditions, target composition, potential overlap with training or tuning data from the same sources, or the annotation procedure and inter-annotator agreement. The number of frames does not equate to the number of independent samples. If the videos originate from only a few observing sessions or instruments, correlations between scenes could substantially influence the reported real-world performance. The authors should report results and confidence intervals calculated at the level of independent observation sequences and clarify how test-set isolation was ensured. The current MOTA/HOTA results on a single, relatively small real-world dataset are better interpreted as preliminary evidence of feasibility than as a reliable validation of transfer to real-world conditions.

Questions for authors
  1. How were the synthetic parameter distributions determined? Specifically, how were the distributions of line-source length, width, brightness, velocity, angle, truncation, PSF, and other relevant parameters selected? Were these distributions fitted to or calibrated against real observations? Could the authors report quantitative differences between the synthetic and real datasets in key geometric, photometric, and motion statistics, and conduct parameter perturbation experiments or ablation studies to assess the robustness of the main conclusions to the simulation assumptions?
  2. What are the sources and degree of independence of the 50 videos in SDTD-Real? How many independent observing sessions, instruments, or data sources do these videos represent? How was separation by observing session or source enforced across training, validation, hyperparameter tuning, and testing? Could the authors provide uncertainty estimates based on per-video statistics, rather than frame-level statistics alone, and report the number of annotators and their inter-annotator agreement?
  3. In Equation (14), the term 1−sj1-s_j is constant for a given detection candidate jj, regardless of which track it is matched to. Under standard one-to-one Hungarian matching, if no explicit dummy nodes, unmatched-assignment costs, or candidate-rejection rules are defined, this term alone cannot change which track the candidate is assigned to. Could the authors clarify how this term specifically suppresses false-positive candidates and how unmatched detections and tracks are handled? Could they also provide comparative experiments that remove this term and the associated matching rules to demonstrate their actual contribution?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.