List of Papers Browse by Subject Areas Author List
Abstract
Existing multi-object tracking methods typically rely on motion models or appearance features for target association. However, in sperm multi-target tracking tasks, the appearance of the targets is highly similar and the motion patterns are complex. It is difficult for both appearance and motion features to exhibit stable tracking performance in targets with different distribution densities simultaneously. Additionally, it is difficult to obtain the motility direction label of sperm targets. This paper proposes Self-supervised Sperm Motility Direction Learning and Density-Adaptive Association for Sperm Tracking (MDL-DA Track). In the sperm detection module, a multi-task learning framework combining self-supervised Sperm Motility Direction Learning (SMDL) and object detection is constructed. Through self-supervised learning of temporal bounding boxes, motion direction information is captured. YOLOX is jointly employed to model both sperm target locations and motion directions. In the sperm tracking module, a motion direction-enhanced Density-adaptive Object Association (DOA) strategy is proposed to handle differences between sparse and dense target distributions. The DOA strategy adaptively selects association rules according to local target density. For targets with sparse distributions, matching is dominated by motion information. For targets with dense distributions, association is achieved by combining motion, appearance, and direction features, thereby enhancing tracking robustness in complex scenarios. Experimental results on the public VISEM-Tracking dataset and Electron Microscopy Sperm dataset show that MDL-DA Track outperforms several state-of-the-art methods in multi-object tracking. On the public dataset, higher order tracking accuracy and multiple object tracking accuracy are improved by 2.5%-3%. Ablation studies further demonstrate that SMDL and DOA effectively reduce erroneous associations in dense scenarios and enhance robustness in sparse scenarios.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3690_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/hjhxy/MICCAI_MDL-DATrack
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZenXia_MDLDA_MICCAI2026,
author = { Zeng, Xiaoyu AND Zhao, Yang AND Wang, Xu AND Fu, Longlong AND Wu, Jiahui AND Pei, Jihong AND Yang, Xuan},
title = { { MDL-DA Track: Self-supervised Sperm Motility Direction Learning and Density-Adaptive Association for Sperm Tracking } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16880},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper draws inspiration from the biological insight that “the direction of sperm movement is consistent with its head-tail orientation” to improve upon traditional tracking algorithms. The authors gained a deep understanding of the clinical challenges faced by conventional tracking methods in sperm tracking—namely, that approaches relying solely on motion and appearance are prone to identity switches—and made targeted improvements accordingly, achieving favorable results compared to baselines. They construct self-supervised signals using short-term frame displacements, thereby avoiding the cost of manual annotation. Furthermore, by integrating local density with feature weights, the method alleviates the problem of trajectory overlap in dense regions.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The strengths of the paper are as follows: (1) Proposed a self-supervised sperm motility direction learning approach based on inter-frame displacement of the sperm, which provides geometry-aware orientation cues that remain stable under dense interactions and abrupt motion, improving association reliability. (2) Proposed a two-stage sperm target association for sparse-region sperm matching and dense-region sperm matching, which further, which enhances density adaptive.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
(1) The algorithm performance in static sperm scenario is not described clearly. The author only mentioned it in the conclusion section. Additional experiments are needed to evaluate its performance on static sperm scenario. (2) It should be clearly explained why the branch weight λ=0 still improves the performance.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
No concern.
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
(1) Fig. 1 should be briefly introduced in the context so readers know its purpose. (2) Paragraph 3 in the Methodology section should be moved after equation (1) to explain the parameters in the formula.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper proposed a self-supervised approach to learning sperm motility direction and density-adaptive association for sperm tracking, which significantly improved tracking performance compared to other SOTA algorithms. Ablation study confirmed the contribution of each module of the approach. However, the paper has some weaknesses, such as the lack of a static sperm scenario and a detailed explanation of the parameters in the equations.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper proposes MDL-DA Track, a two-stage multi-object tracking framework tailored for sperm tracking that augments a YOLOX detector with a self-supervised direction branch (SMDL) and introduces a density-adaptive association (DOA) strategy. SMDL predicts per-detection orientation (sinθ, cosθ) using pseudo-labels from short-term displacement, while DOA switches between motion-dominant sparse matching and a parallel motion/appearance association in dense regions, both enhanced by the learned direction cue. On VISEM-Tracking and an Electron Microscopy Sperm dataset, the method reports consistent improvements over strong baselines (e.g., TOPICTrack, OCSort), with 2.5–3% gains in HOTA/MOTA and stronger identity preservation metrics.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
A. Technical novelty and innovation
-
The addition of a lightweight, geometry-aware direction head within YOLOX trained via displacement-derived pseudo-labels is a novel approach, where a domain-informed extension that decouples orientation from displacement magnitude.
-
The density-adaptive association strategy is practical and aligns with the unique sparsity/density variability of sperm videos. In which authors integrate learned direction consistently into both sparse and dense association costs is a coherent design.
B. Experimental validation
-
Fair comparisons are conducted by using the same detector across methods; ablations isolate the impact of DOA and SMDL, and a hyperparameter sweep for \lambda is provided.
-
Results are reported as mean +- standard deviation across splits, improving statistical credibility relative to single-split reporting.
C. Clarity of presentation
-
The methodological flow is clear: detector with SMDL -> density estimation -> two-stage association with well-specified cost terms and components.
-
The choice to represent direction via (sinθ, cosθ) and the constraint to positive samples with sufficient displacement are well-motivated.
-
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
It’s still unclear to reviewers the following:
-
First, “Self-supervised” direction learning indeed relies on ground-truth inter-frame associations (to compute displacement labels), which are weak/self-labeled rather than unsupervised; the dependence on annotated tracks thus should be clarified.
-
Several elements referenced in the method, e.g., “motion reliability,” trajectory-level direction aggregation, are not formally defined, leaving ambiguity in implementation details critical for reproducibility.
-
More ablation studies are required to provide (to the appendix should be fine). They include: (i) breakdown of performance across sparse vs. dense; (ii) comparisons to segmentation-based or microscopy-specific trackers (e.g., classical cell/sperm tracking pipelines), which are currently absent.
- Other points:
- The current figure method needs to be polished for further highlight adaptations to the sperm tracking. For instance, the “displacement computation” should be replaced by the optical flow vectors demonstrating displacement, etc.
-
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Both the method contributions and experiments are significant and well-motivated. The paper can be further improved by enhancing the visualization of the method figure and by providing details, including additional ablation studies on parameter sensitivity.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
I have read the rebuttal, which addressed my concerns. I decided to keep my current positive evaluation.
Review #3
- Please describe the contribution of the paper
The paper presents MDL-DA Track, a sperm multi-object tracking framework for microscopic videos. The authors claim two main contributions: 1- Self-supervised Sperm Motility Direction Learning (SMDL), which uses a YOLOX-based detector with an additional direction prediction branch trained using pseudo-labels made from temporal bounding box movement 2- Density-Adaptive Object Association (DOA), which changes the association strategy based on local target density, relying more on motion movements in sparse regions and combining motion, appearance, and direction vector in dense regions. The reported experiments on VISEM-Tracking and the Electron Microscopy Sperm dataset show improvements over several existing trackers in HOTA, MOTA, IDF1, and AssA.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
The paper is generally well structured and easy to follow. The motivation, method, and experiments are presented in a coherent way, and the two main components of the approach, SMDL and DOA, are described clearly enough that the overall pipeline is understandable.
Another strength is that the paper addresses a clearly defined problem. The authors identify that, in sperm tracking, appearance cues are weak because the targets are highly visually similar, while motion cues become less reliable in dense and crowded regions. The proposed method is at least well aligned with these stated challenges, with SMDL intended to provide an additional direction cue and DOA intended to adapt association behavior to local density.
The paper also has practical application relevance. Accurate sperm tracking is an important component of computer-aided sperm analysis systems, and improvements in association-related tracking metrics may be useful for downstream motility analysis. While the paper does not demonstrate clinical translation directly, the application itself is meaningful.
A further positive aspect is that the paper includes ablation and sensitivity analyses rather than only a single comparison table. This makes the experimental section more complete and helps show the individual effect of the proposed components.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The novelty appears limited. The two core ideas are direction-aware association and adaptive use of different association cues depending on the tracking situation, but both themes already exist in prior MOT literature. OC-SORT explicitly incorporates direction consistency into association to improve robustness under non-linear motion and occlusion, while TOPIC/TOPICTrack already proposes a parallel motion/appearance association framework with conflict resolution and adaptive cue selection. In that context, the paper is more of an adaptation of existing MOT ideas to sperm tracking than a new tracking formulation.
The experimental comparison is a notable weakness of the paper. Although the paper includes a few newer baselines, the comparison in Table 1 remains dominated by older generic MOT methods, even though the paper’s main claims focus on sperm tracking and robustness in dense, crowded scenes. In particular, the paper does not compare against relevant papers such as Sperm YOLOv8E-TrackEVD (https://www.proquest.com/docview/3067437403), YOLOv4 with improved DeepSORT for sperm detection and tracking (https://www.nature.com/articles/s41598-025-26465-4), and an IMM-based multi-sperm dynamic tracking algorithm (https://www.sciencedirect.com/science/article/pii/S1746809425006834). Also, the dense-scene claim would be more convincing if the paper were compared against stronger crowded MOT baselines such as ConfTrack (https://openaccess.thecvf.com/content/WACV2024/html/Jung_ConfTrack_Kalman_Filter-Based_Multi-Person_Tracking_by_Utilizing_Confidence_Score_of_WACV_2024_paper.html), TrackTrack (https://cvpr.thecvf.com/virtual/2025/poster/35174), and Hybrid-SORT (https://ojs.aaai.org/index.php/AAAI/article/view/28471), all of which address occlusion, identity preservation, or association in crowded settings. Even if not all of these methods are directly optimized for microscopy, the absence of stronger recent sperm-specific and dense-scene comparators makes the empirical case less convincing.
Finally, some of the most important choices in DOA are not analyzed. The paper includes an ablation study and a sensitivity analysis of the direction-loss weight, which are useful, but it provides much less evidence on sensitivity to the density threshold, adaptive neighborhood radius, and association weights.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The paper is clearly written, and the overall idea is easy to follow. The problem is well motivated, and the reported results show consistent gains over the baselines. However, my main concern is that the methodological novelty appears limited relative to prior MOT literature, since the core ideas of direction-aware association and adaptive use of different matching cues are already present in prior works such as OC-SORT and TOPIC/TOPICTrack. As a result, the method feels more like a task-specific adaptation to sperm tracking than a substantially new tracking formulation. In addition, the experimental case is not fully convincing. Most of the compared methods are either generic MOT trackers or older baselines, and I would have liked to see comparisons against more recent general MOT and sperm-specific tracking methods. I would also like to see more ablation of the DOA design choices, especially the density-related parameters and the association weighting.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the reviews and the rebuttal, I acknowledge the effort by the authors to differentiate their contribution from the previous. I think it should be made even clearer this differentiation in the camera-ready, and assuming that the authors will do a good job with that, I am willing to accept the paper.
Author Feedback
We thank reviewers for recognizing novel (R2), meaningful, easy to follow, clearly written, experimental section complete (R3), well-motivated (R2, R3), clearly defined problem (R1, R3), significant improvement (R1, R2).
Innovation (@R3) Sperm tracking suffers from uneven distribution, target overlap in dense scenes, and illumination interference in sparse scenes, making motion direction information indispensable. Given that sperm head and tail orientation generally align with its moving direction, we propose MDL-DA Track. Its core innovations are: 1) SMDL module generates pseudo motion direction labels from target positions to supervise the lightweight direction branch, enabling the detector to learn motion direction cues from image features. 2) DOA strategy estimates local target density and adaptively adjusts the association weights of IoU, appearance and motion direction features. Ours Vs. OC-SORT Although both methods use directional information, they differ fundamentally in motivation and information source. OC-SORT infers trajectory direction from historical trajectories and inter-frame position changes. Therefore, it relies on trajectory continuity and motion estimation quality, which may be affected in dense sperm scenes with frequent occlusion, interaction, and irregular motion. In contrast, our SMDL module learns morphologically aware directional representations from appearance. It uses inter-frame displacement as supervision and directly infers motion-related direction cues from current frame features without relying on accumulated trajectory history. Our method further integrates directional learning into both detection and association via joint optimization, forming a specialized framework for sperm tracking. Ours Vs. TOPICTrack Our innovation lies not in redeveloping parallel association, but in designing a density-adaptive mechanism. TOPICTrack employs a unified association strategy for all targets, selecting IoU or appearance cues based on the degree of motion change. Sperm tracking features uneven in-frame density, where association cues vary substantially between sparse and dense regions — a scenario ignored by TOPICTrack. Our DOA adopts different strategies for different density regions: motion matching is mainly used in sparse regions, while fusion of motion, appearance and direction is used in dense regions. In addition, the mechanism is structurally independent and compatible with other MOT frameworks.
Meaning of Self-Supervision (@R2) The term “self-supervised” here means that our direction learning does not require additional manual orientation annotations, such as sperm head-tail orientation or motion direction labels. We acknowledge that this is not strict self-supervision and will clarify this definition in the revised manuscript.
Ablation on branch weight λ (@R1): Table 1 first row denotes the baseline without our DOA. In contrast, when λ=0, the model still retains the full DOA workflow, and the direction branch remains active during forward propagation and modulates shared features. This enables the backbone to capture richer geometric structural information via structural inductive bias. We will add these explanations to the revised manuscript.
Other ablation experiments (@R1, R2, R3): We have already conducted more detailed ablation experiments, but they were not included in the original manuscript due to page limits. We will add more ablation results in the revised version if space allows.
Comparison Experiments (@R2, R3): We compare with representative MOT methods including ByteTrack and BoT-SORT, which address dense tracking challenges via detection confidence, appearance modeling, motion estimation and temporal memory. We agree to add more domain-specific methods. Other details (@R1, R2): We will revise the descriptions of key concepts, unify and standardize notations, and provide clearer definitions for core symbols. Code: We will release the source code publicly upon paper acceptance.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
The paper proposes a self-supervised framework for sperm tracking with direction learning and density-adaptive association. The reviewers’ scores are 5, 4, and 3, indicating mixed opinions. The reviewers agree that the method is well motivated and demonstrates promising improvements over existing approaches. However, concerns remain regarding the level of methodological novelty, the clarity of certain design choices, and the completeness of experimental validation. In particular, the novelty should also be clarified with respect to prior cell tracking studies that jointly model detection and motion estimation. The AC therefore recommends inviting the paper for rebuttal.
The authors are encouraged to carefully address the reviewers’ comments in the rebuttal, particularly regarding the novelty with respect to existing tracking methods, additional analysis of the proposed components, and the completeness of experimental evaluation and comparisons.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper proposes MDL-DA Track, a sperm multi-object tracking framework for microscopy videos. It combines a sperm motility direction learning branch trained from inter-frame displacement pseudo-labels with a density-adaptive association strategy that changes the use of motion, appearance, and direction cues between sparse and dense regions. The problem is practically relevant, and the method reports consistent improvements on VISEM-Tracking and an Electron Microscopy Sperm dataset.
The initial reviews were mixed but overall close to positive, with scores of 5 / 4 / 3.Reviewers appreciated the clear motivation, domain-informed use of direction cues, density-adaptive association design, and consistent empirical gains. The main concerns were limited novelty relative to prior MOT methods such as OC-SORT and TOPIC/TOPICTrack, incomplete comparison with recent sperm-specific or crowded-scene trackers, insufficient analysis of static sperm cases and DOA design choices, ambiguity in the term “self-supervised,” and missing implementation or notation details.
The rebuttal addressed several of these concerns. The authors clarified how their direction learning differs from trajectory-history-based direction modeling in OC-SORT, and how their density-adaptive association differs from the cue-selection strategy in TOPICTrack. They also clarified that “self-supervised” means no additional manual orientation labels rather than fully unsupervised learning, explained the lambda=0 ablation, and committed to clearer notation, definitions, additional ablations where space allows, more domain-specific comparisons, and code release. These responses were sufficient for the reviewers to converge to a positive recommendation.
Some limitations remain. The contribution is still best viewed as a well-designed, domain-specific adaptation and integration of existing MOT ideas rather than a fundamentally new tracking formulation. The comparison to recent sperm-specific and dense-scene trackers remains incomplete, and the sensitivity of DOA parameters such as density thresholds, neighborhood radius, and association weights should be better clarified. The camera-ready version should also clearly define the self-supervised setting and provide the missing implementation details.
After rebuttal, the final recommendation is A / A / A. Reviewer 2 kept a positive assessment after finding that the rebuttal addressed the main concerns, and Reviewer 3 moved from Weak Reject to Accept based on the authors’ effort to better differentiate the contribution from prior work. The score improvement mainly reflects improved confidence in the novelty framing, domain-specific motivation, and planned clarification of reproducibility details.
I recommend acceptance. The post-rebuttal reviewer consensus is positive, and the paper addresses a meaningful microscopy tracking problem with a coherent domain-informed method and consistent empirical evidence. While the remaining concerns on novelty, comparator coverage, and DOA parameter analysis should be noted, they are not sufficient reasons for rejection and can be addressed in the final version.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
After rebuttal two reviewers are happy with the response. And the other reviewer did not provide post-rebuttal decision. But his pre-rebuttal recommendation is Accept. So all three reviewers are positive. Overall this is a well-written paper with good novelty and experimental verification. So the final recommendation is Accept.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
Reviewers’ concerns were solved after the rebuttal. After reading the paper, rebuttal and the reviews after the rebuttal, the AC recommends accepting this paper which should be revised based on the three reviewers’ feedback.