2.1. Deep learning in biology research

Thanks to its significant advantages in data-driven inference, deep learning has been widely applied in biological research, greatly accelerating the progress of modern biology. At the molecular and genomic level, DeWeirdt et al. [12] combined language models and machine learning algorithms to develop a predictive model of prokaryotic immune systems, capable of accurate predictions on previously unseen data. Charlier et al. [13] compared different neural network architectures for CRISPR-Cas9 off-target prediction and proposed a similarity-based transfer learning approach. Afting et al. [14] utilize deep learning to predict the differentiation path and resulting tissues in retinal organoids well before they become visually discernible. Zhang et al. [15] developed a framework to infer cell line-specific molecular features from transcriptional regulation networks, integrating three modalities of input data, widely used in genome annotation. At the organismal level, deep learning has been applied to behavior tracking and biomechanics. For example, Ulrich et al. [16] employed pose estimation to track movements of worker bees and queens, combining machine learning to extract behavioral metrics and establish a bee behavior tracking system. Melis et al. [17] investigated insect wing biomechanics, using convolutional neural networks to predict the relationship between controlling muscle activity and wing motion. Han et al. [18] collected video recordings of two mice, applied multi-stage artificial neural networks to identify individuals and their postures, and generated three-dimensional trajectories for unsupervised multi-animal behavior recognition.

In silkworm research, deep learning has been applied to various tasks. Guo et al. [6] developed a posture correction algorithm for pupae, improving sex identification accuracy. He et al. [19] combined near-infrared spectroscopy with deep learning to identify developmental morphology, sex, and variety. Shi et al. [20] proposed an enhanced object detection method for individual recognition. These studies illustrate the applicability of deep learning in silkworm research, providing a reference for MOT and related downstream tasks.

2.2. Current studies in MOT

Multiple object tracking (MOT) is a significant subfield of visual deep learning with applications in autonomous driving, robotics, and smart agriculture. MOT methods are typically classified into Tracking by Detection (TBD) and End-to-End approaches [11]. TBD methods, which offer significant advantages in tracking accuracy, have gained considerable attention in the industry. SORT and DeepSORT [10] are pioneering methods in this field, using Kalman filters to update trajectory states and the Hungarian algorithm to associate deep feature vectors between trackers and detected objects, resulting in substantial improvements in tracking performance. Li et al. [21] proposed a real-time detection and counting method for wheat ears based on DeepSORT. Zhang et al. [22] integrated convolutional auto-encoders into DeepSORT to enhance appearance feature extraction, introducing a tracking method for young honey peach fruits. Cao et al. [23] developed a dynamic counting method for sheep using DeepSORT. Meng et al. [24] proposed a real-time tracking method for cherry tomato count, utilizing a lightweight network to extract deep features and optimize matching efficiency. Tu et al. [25] introduced a passion fruit yield prediction method based on OC-SORT and YOLOv8. Optimizing data association methods remains a central focus in MOT research. Li et al. [26] described a multi-scale fusion network for object feature matching, employing similarity matrix loss to calculate associations between objects in different frames. Zhang et al. [27] used a spatiotemporal association model to extract historical trajectory correlations and employed graph neural networks to fuse appearance and motion information. Wang et al. [28] proposed a joint detection and matching method utilizing neighboring frame position prediction and inter-frame relationships to predict object positions in the next frame. Liang et al. [29] introduced a tracking algorithm for aerial vehicles, relying on a Gaussian distance module and an observation-centered Kalman filter. Muzaddid and Beksi et al. [30] proposed a tracking method based on linear relationships between adjacent trajectory positions, using dense optical flow and particle filtering to guide each tracker.

These studies highlight the broad applications of MOT and provide valuable references for silkworm tracking. Building upon this prior work, we optimize data matching strategies and propose a MOT method specifically tailored to silkworms.

3.1. Silkworm MOT dataset

In this study, silkworm videos were collected under real-world conditions, preprocessed, and used to construct a MOT dataset. Data acquisition was conducted on September 10, 2023, at the Sericultural Research Institute of the Sichuan Academy of Agricultural Sciences, Sichuan Province, China. A mobile camera (Moker High-Speed USB 3.0; resolution: 2592 × 1944; frame rate: 25 FPS) was mounted vertically above the rearing boxes to capture video data. The environment was illuminated by incandescent lamps, with the temperature maintained at 29 °C and relative humidity at 80%.

In large-scale silkworm farming, high-density rearing often results in frequent occlusions caused by overlapping mulberry leaves or individual blockage, posing significant challenges for data collection and annotation. Given that silkworm MOT is intended for biological applications such as strain breeding, behavior analysis, and disease prediction, this study focused on a breeding environment with relatively low population density. After a period of feeding, silkworms typically emerge from beneath the mulberry leaves and consume most of them, thereby substantially reducing occlusion.

Two videos were recorded in this breeding environment, each with a duration of approximately 1.5 hours. Due to the slow crawling speed of silkworms, frames were sampled at 3-second intervals, resulting in two datasets: MST-50 (multi-silkworm tracking with approximately 50 individuals across 1,000 frames) and MST-100 (approximately 100 individuals across 1,200 frames). The spatial distribution of silkworms in the dataset is illustrated in Fig 1. Following the MOT17 dataset construction methodology [31], the position of each silkworm was manually annotated across all frames using LabelImg, yielding over 50,000 and 120,000 individual annotations for MST-50 and MST-100, respectively. These annotations were extracted from the generated XML files and converted into text format to form the final MOT dataset.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Schematic images of silkworm MOT dataset.

(a) Frame 1 of MST-50. (b) Frame 300 of MST-50. (c) Frame 1 of MST-100 (d) Frame500 of MST-100.

https://doi.org/10.1371/journal.pcbi.1014410.g001

Table 1 presents a sample of the annotation file, where each line corresponds to a silkworm instance in a single frame, including the frame index, identity ID, bounding box center coordinates, and bounding box dimensions.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Schematic of the annotation file.

https://doi.org/10.1371/journal.pcbi.1014410.t001

3.2. WormSORT for silkworm MOT

To enable precise tracking of individual silkworms in a breeding environment, we propose a multiple object tracking (MOT) method based on a tracking-by-detection (TBD) framework, termed WormSORT. As illustrated in Fig 2, WormSORT consists of three main components: a detector (the pretrained YOLOv10s), a feature extraction model (BoT-S50 pretrained on the DukeMTMC-reID dataset), and a data association module. The detector is employed to localize silkworms in each frame, while BoT-S50 extracts appearance features of the detected targets. Data association is first performed using the Intersection over Union (IoU) metric, followed by feature-based matching to improve robustness and accuracy. To further enhance the matching performance, we introduce IoU padding and feature padding strategies, which allow high-confidence unmatched trackers and detections to remain in the association process.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Structure of WormSORT.

https://doi.org/10.1371/journal.pcbi.1014410.g002

3.2.1. Algorithm flow of WormSORT.

When WormSORT is used to track silkworm trajectories across consecutive frames, YOLOv10s [32] first localizes the silkworms, and the pre-trained BoT-S50 [33] extracts their depth features. A detection set is generated: , where each detection contains position information, a depth feature vector, and a detection confidence score. Since the initial tracking set is empty, no association is performed in the first frame. Instead, the detection set is directly initialized as the tracking set: , where each tracked individuals is assigned an identity ID, position, state vector, and covariance matrix , which are subsequently updated using a Kalman filter In addition, each track maintains its depth feature vectors and the number of confirmed frames.

In the second frame, the same detection and feature extraction process is repeated, producing a new detection set . Since the trajectories in have not yet been confirmed, detections in are initialized as new trajectories and merged with to form . The Kalman filter is then applied to predict the likely positions of each trajectory in the next frame.

From the third frame onward, detections and predicted positions are used to perform association. First, the IoU values between and are calculated to form a cost matrix. The Hungarian algorithm is applied to obtain the first-round associations. Unmatched detections are further compared to existing tracks using depth features. The cosine distance between detection and track feature vectors forms a new cost matrix, which is again solved using the Hungarian algorithm.

Unmatched detections are initialized a s new tracks and added to the tracking set. Unmatched tracks are marked as tentative; if a tentative track remains unconfirmed for more than a predefined number of frames , it is removed. This iterative process yields the complete crawling trajectories of silkworms. A detailed stepwise flow of WormSORT is provided in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Algorithmic flow of WormSORT.

https://doi.org/10.1371/journal.pcbi.1014410.t002

3.2.2. Data association of WormSORT.

The data association procedure in WormSORT is conceptually similar to DeepSORT, which relies on IoU and feature-based matching. However, unlike DeepSORT, WormSORT first performs IoU-based matching, followed by feature-based matching. This design leverages the observation that silkworm positions and states exhibit temporal continuity, thereby reducing the likelihood of mismatches.

The IoU between a detection and a predicted track is defined as:

(1)

The corresponding cost matrix is expressed as:

(2)

where represents the position of the -th detection in the frame , and denotes the predicted position of the -th tracker after Kalman filtering.

The Hungarian algorithm was used to explore the optimal matching solution by solving the following optimization problem:

(3)

where represents the matching set, is the detection set, and is the tracker set.

For unmatched detections and trackers, the cosine distance between their depth feature vectors is calculated to form the cost matrix:

(4)

where denotes the depth feature vector of the -th detection, and denotes the depth features vector of the -th tracker.

The Hungarian algorithm is then applied to minimize the feature-based cost matrix:

(5)

The final matching result is obtained by combining both stages:

(6)

Thus, the data association strategy in WormSORT integrates the efficiency of IoU-based matching with the accuracy of depth feature matching, thereby ensuring both robustness and tracking efficiency, which is particularly effective given the relatively stable positions of silkworms.

3.2.3. Candidate input padding mechanism.

Multi-object tracking methods based on object detection rely heavily on the accuracy of the detector. However, in silkworm tracking, the close proximity of individuals makes the system highly sensitive to minor detection failures, which can severely reduce the accuracy of data association. Moreover, the high visual similarity among silkworms often causes overlapping bounding boxes, diminishing the discriminative power of depth features.

To address these challenges, WormSORT introduces two complementary padding strategies: IoU-padding, which exploits the spatial continuity of trajectories, and feature-padding, which leverages the similarity of depth features. Together, these mechanisms compensate for missed detections and reduce mismatches by augmenting the candidate set before association.

1. (1). IoU padding

Prior to IoU-based matching, the mechanism considers not only the trajectories confirmed in the previous frame but also unconfirmed trajectories and detections carried over from earlier frames. Specifically, if the geometric position of an unmatched tracker in the previous frame and the center point of a detection box within the following ten frames satisfy the Euclidean distance condition, the trajectory is retained for further consideration:

(7)

where denotes the predefined motion-range threshold, set to 5. represents the set of unmatched trackers from the previous frame.

Through this augmentation, unconfirmed yet high-quality trackers from the past ten frames are reintroduced into the matching set. Such trackers may have been excluded because two silkworms were positioned in close proximity, leading them to appear as secondary costs in the assignment matrix, or due to missed detections that prevented their confirmation in earlier frames. Accordingly, the candidate tracker set is defined as:

(8)

where denotes the set of confirmed trajectories from the previous frame.

This strategy constitutes the IoU padding mechanism, ensuring that spatially consistent yet unconfirmed trackers remain available for matching.

1. (2). Feature padding

During the deep feature matching stage, supplementary matching is performed only for detection boxes that failed in IoU-based matching. In addition, trajectories in whose cosine similarities to certain detection boxes within the most recent ten frames satisfy the following condition are also included:

(9)

where denotes the set of trajectories unconfirmed by IoU matching, and denotes the predefined cosine similarity threshold, set to 0.8.

Unmatched trajectories that satisfy this condition are incorporated into the feature candidate set:

(10)

This strategy constitutes the feature padding mechanism, ensuring that high-confidence trajectories based on feature similarity are preserved.

Taken together, WormSORT employs both IoU-padding and feature-padding mechanisms, which jointly compensate for detection failures and mismatches. By fully exploiting both spatial continuity and feature similarity, these mechanisms enhance robustness and ensure continuity in silkworm trajectory tracking.

3.2.4. Detector training and appearance feature extraction.

WormSORT utilizes YOLOv10s for its favorable balance between efficiency and detection accuracy. Architecturally, it follows the standard YOLO paradigm, consisting of a backbone, a neck, and three detection heads, and performs object localization and classification via fully convolutional operations. YOLOv10 further introduces a consistent dual-assignment strategy for NMS-free training to mitigate redundant predictions. Additionally, a lightweight classification head, spatial–channel decoupled downsampling, and rank-guided blocks are employed to reduce computational redundancy. The integration of large-kernel convolutions and an efficient partial self-attention module further enhances representational capacity with minimal overhead.

To adapt the detector to the silkworm tracking scenario, YOLOv10s was trained on a dedicated dataset comprising 1,526 images with 83,698 annotated instances, averaging over 50 silkworms per image. Representative training samples are shown in Fig 3, demonstrating significant variations in object scale and spatial distribution. The dataset was split into training and validation sets at a ratio of 7: 3. During training, Mosaic augmentation was applied, with an input resolution of 720 × 720 pixels. Optimization was performed using the Adam optimizer with the CIoU loss for 300 epochs, an initial learning rate of 0.001, and a cosine decay schedule.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. Image examples of object detection dataset.

https://doi.org/10.1371/journal.pcbi.1014410.g003

The trained detector achieved a recall of 97.37%, a precision of 97.84%, and an mAP@50 of 98.63% on the MST-50 dataset. Despite these competitive detection results, the tracking performance remained suboptimal when directly applied to MOT. To address this issue, 500 additional images were sampled from MST-50 at regular intervals and incorporated into the training set for further fine-tuning. The refined detector achieved a recall of 99.79%, a precision of 99.81%, and an mAP@50 of 99.50%, and was subsequently used for experiments on MST-50. The impact of the number of sampled images on tracking performance in MST-100 is further analyzed in Section 4.2.

For appearance modeling, a BoT-S50 model with a ResNet-50 backbone, implemented in the FastReID framework and pretrained on the DukeMTMC-reID dataset, is used to extract discriminative appearance features, which aid in robust data association during tracking. Based on the detector's predictions, the position of each silkworm is determined, and its bounding box is extracted. The image is then cropped, resized to 256 × 128 pixels, and passed through the model to extract a 1024-dimensional feature vector, which serves as the appearance feature.

3.3. Experimental setting and evaluation metrics

Experiments were conducted on a DELL 5820 workstation equipped with an Intel Core i7-9800X CPU and NVIDIA RTX 2080Ti GPUs (11 GB memory), using CUDA 10.0. The operating system was Windows 10 Professional (64-bit). All models were implemented in Python 3.7 using the PyTorch 1.7 framework, with auxiliary libraries including NumPy, Keras, and OpenCV.

To evaluate the impact of the detector on tracking performance on the MST-100 dataset (Section 4.2), two training strategies were compared: training from scratch and transfer learning. Additional training images were uniformly sampled from MST-100 and incrementally incorporated into the original training set. For training from scratch, the same hyperparameters as those used for the MST-50 detector were adopted. For transfer learning, training was initialized with 100 epochs, and the number of epochs was increased by 10 with each increment of added training images. In each transfer learning stage, training was initialized from the baseline detector weights, without sequential fine-tuning across stages. After training, the detector was evaluated on MST using recall, precision, F1-score, and mAP@50.

For MOT evaluation, the detection confidence threshold was set to 0.3, the cosine similarity threshold for feature matching to 0.4, and the intersection-over-union (IoU) threshold to 0.7. The maximum retention length of unconfirmed trajectories was set to 30 frames.

Tracking performance was evaluated using five widely adopted metrics: Multi-Object Tracking Accuracy (MOTA), IDF1 score (IDF1), Association Accuracy (AssA), Detection Accuracy (DetA), and Higher-Order Tracking Accuracy (HOTA). Their formulations and corresponding definitions are as follows.:

(11)

where denotes the number of ground-truth objects at time , represents the number of false negatives at time , denotes the number of false positives at time , corresponds to the number of identity switches at time .

(12)

where denotes identity true positives, denotes identity false positives, denotes identity false negatives.

(13)

where denotes the number of correctly matched target association pairs, denotes the number of incorrectly matched pairs, and denotes the number of unmatched pairs.

(14)

where denotes the number of correctly detected targets, denotes the number of falsely detected targets, and denotes the number of undetected targets.

(15)

4.1. Tracking experiments on the MST-50 dataset

To evaluate the performance of the proposed WormSORT, we compared it against several state-of-the-art multi-object tracking methods, including DeepSORT, StrongSORT [34], ByteTrack [35], OCSORT [36] and BotSORT [37]. All experiments were conducted under identical settings.

Fig 4 presents the tracking results on the MST-50 dataset. WormSORT achieved 96.37% in MOTA, 93.65% in IDF1, 91.33% in HOTA, and 89.85% in AssA, demonstrating substantial improvements in identity consistency, association accuracy, and tracking stability. A slight decline in DetA (92.93%) was observed, likely due to the reliance on IoU-based matching and Kalman filtering, which may be less effective for nonlinear head-swinging motions of silkworms.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Comparison results with related methods on the MST-50 dataset.

https://doi.org/10.1371/journal.pcbi.1014410.g004

Compared to StrongSORT, WormSORT achieves a significant performance improvement, especially in IDF1 and AssA. This enhancement can be attributed to WormSORT’s strategy of prioritizing IoU-based matching followed by deep feature matching, combined with IoU padding and feature padding, which ensures better track association, particularly under partial occlusions and brief detection misses.

ByteTrack, which relies solely on IoU-based detection, achieved lower HOTA (77.98%) and AssA (75.97%), highlighting the limitations of detection-only association for silkworms’ motion. OCSORT, designed to mitigate occlusions using virtual trajectories, achieved high DetA (94.12%) but suffered in IDF1 (76.58%) and AssA (67.62%) due to its inability to capture nonlinear silkworm motions, indicating that virtual trajectory strategies alone are insufficient for this scenario.

BotSORT, which integrates both IoU and deep feature matching, achieved strong performance in IDF1 (89.68%) and HOTA (81.18%), but still underperformed compared to WormSORT, demonstrating the benefit of the proposed padding strategies and motion assumptions in accurately maintaining identities over time.

Overall, WormSORT achieves the best tracking performance across most evaluation metrics, effectively balancing accurate detection association and robust identity preservation, thereby demonstrating its suitability for silkworm MOT in this dataset.

Fig 5 presents a qualitative comparison of six methods at a late tracking stage (frame 800), highlighting identity switch behavior. The maximum assigned identity (ID) in each method is marked with red ellipses. DeepSORT reaches a maximum ID of 213, followed by StrongSORT (80), ByteTrack (96), OCSORT (115), and BoTSORT (106). In contrast, WormSORT yields a substantially lower maximum ID of 67. Although occasional short-lived detections (e.g., low-quality or transient targets) may lead to slight increases in ID count, the visualization indicates that WormSORT effectively suppresses identity switches and maintains superior tracking stability.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. Tracking comparison of six methods at Frame 800 in MST-50 dataset.

https://doi.org/10.1371/journal.pcbi.1014410.g005

In summary, WormSORT demonstrates clear advantages in MOTA, IDF1, HOTA, and AssA, while maintaining competitive performance in DetA. These results collectively confirm that WormSORT achieves the best overall performance for multiple silkworm tracking.

4.2. Tracking experiments on the MST-100 dataset

The detector plays a critical role in multiple object tracking, especially in scenarios with complex motion and frequent occlusion such as silkworm tracking. In practice, directly applying a detector trained on general datasets often leads to suboptimal tracking performance due to domain differences. To better understand this effect, we construct task-specific training sets by sampling images from the MST-100 dataset and retrain the detector under different data scales and training strategies (Transfer learning and training from scratch). The resulting detectors are then integrated with WormSORT for tracking evaluation.

Fig 6 presents the tracking results of WormSORT on the MST-100 dataset with different detector configurations, demonstrating that the detector training strategy has a significant influence on tracking performance. For the baseline model, where no images from MST-100 were included in the training set, but were used entirely as the test set, the detector achieved a Recall of 93.78%, a Precision of 98.21%, an F1-score of 0.96, and an mAP@50 of 98.02%. These results are highly competitive in object detection, indicating strong detection capability. However, when directly applied to silkworms tracking, the performance is relatively limited, with IDF1 and AssA only reaching 49.75% and 32.47%, respectively.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. Tracking performance with different detector training strategies on the MST-100 dataset.

https://doi.org/10.1371/journal.pcbi.1014410.g006

Based on the baseline model, transfer learning not only reduces training cost but also enables the detector to quickly adapt to the distribution of silkworms in the MST-100 dataset, leading to improved detection performance. When 100 images are sampled for training, Recall increases from 93.78% to 94.81%, and mAP@50 improves from 98.02% to 98.47%, while Precision decreases from 98.21% to 96.61%. Despite this, all tracking metrics show noticeable improvements.

As the number of training images increases, the performance exhibits certain fluctuations. When the dataset size increases from 200 to 300 images, the Recall slightly decreases, while tracking performance continues to improve. A similar decrease in Recall is observed when increasing the dataset size from 500 to 600 images. These observations suggest that improvements in detection metrics do not always directly translate into better tracking performance, and the relationship between detection and tracking is not strictly monotonic.

When increasing the dataset size from 300 to 400 images, both Recall and Precision improve, but tracking performance declines. For example, IDF1 decreases from 61.82% to 58.04%, and HOTA decreases from 60.52% to 56.79%. This indicates that better detection accuracy does not necessarily guarantee improved data association quality. Among all transfer learning configurations, the detector trained with 500 images achieves the best overall performance in both detection and tracking metrics. Overall, transfer learning improves tracking performance compared to the baseline, but increasing the training data does not always lead to consistent gains, and performance fluctuations can be observed.

Training from scratch generally requires more computational resources and training time, but produces more stable trends in this task. When 100 images are added to the training set, Recall increases from 93.78% to 96.13%, while Precision decreases from 98.21% to 97.85%. Both F1-score and mAP@50 improve. However, in tracking evaluation, although MOTA and DetA increase, IDF1, HOTA, and AssA decrease significantly, indicating that the limited amount of data is insufficient for learning robust representations.

As more images are added, the detector performance improves steadily. Except for a slight decrease in Precision at 500 and 600 images, Recall shows a consistent upward trend, while F1-score and mAP@50 either improve or remain stable. In terms of tracking performance, MOTA and DetA increase steadily with the number of training images and reach their highest values at 600 images, with 94.93% and 85.22%, respectively. IDF1 reaches its peak at 500 images (64.32%) and shows a slight decrease at 600 images, which may be attributed to minor noise introduced with increased detections. For HOTA and AssA, a slight decrease is observed when increasing the dataset size from 400 to 500 images, possibly due to the reduction in Precision. Other metrics generally show an increasing trend, and both HOTA and AssA achieve their best values at 600 images, reaching 64.81% and 49.37%, respectively.

The results demonstrate that the performance of the detector has a substantial impact on tracking, including both the training data and the training strategy. Otherwise, even if competitive detection metrics such as Precision or mAP are achieved, the tracking performance may still be unsatisfactory. From the MST-100 experiments, transfer learning tends to improve detection metrics, but such improvements do not always lead to better tracking performance. In contrast, training from scratch, although more computationally expensive, generally yields better overall tracking performance. Therefore, sampling images from the target scenario and training the detector from scratch is a key step for achieving reliable tracking results in the silkworm tracking task.

Under the condition that Precision is maintained at a reasonable level, Recall plays an important role in improving multi-object tracking performance. Higher Recall provides more candidate detections, which helps maintain trajectory continuity. Missing detections, on the other hand, are more likely to cause identity switches, leading to lower IDF1 and AssA scores.

Compared to detection, improving tracking performance remains more challenging, especially in complex and long-term scenarios. This is reflected in the relatively lower values of IDF1, HOTA, and AssA. The main reason is that tracking operates under dynamic conditions, where errors accumulate over time. Missing detections or slight noise can lead to track fragmentation, and once an identity is lost or mismatched, it may trigger a cascade of tracking errors, significantly degrading reliability, even when the detector provides largely correct results.

In practical scenarios, occlusion and head movement of silkworms may further complicate data association. From a biological perspective, avoiding overly crowded conditions and excessively long tracking durations can help better capture the crawling trajectories of silkworms and ensure the reliability of subsequent analyses.

After analyzing the impact of different detector training strategies on tracking performance, the detector that achieved the best results with WormSORT was selected for further comparison experiments. The tracking results on the MST-100 dataset are shown in Fig 7. Compared with other tracking methods, WormSORT achieves the best performance in terms of MOTA, IDF1, HOTA, and AssA, while ranking second in DetA, only slightly lower than OCSORT. These results indicate that, compared with existing methods, the data association strategy adopted in WormSORT remains effective in complex scenarios. Overall, WormSORT demonstrates good adaptability to the silkworm tracking task, although there is still room for improvement in certain metrics.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. Comparison results with related methods on the MST-100 dataset.

https://doi.org/10.1371/journal.pcbi.1014410.g007

Fig 8 illustrates representative tracking results of WormSORT on the MST-100 dataset. Red ellipses highlight densely populated regions where identity switches are more likely to occur. At Frame 4, the IDs in this region range approximately from 60 to 90. By Frame 200, the ID range increases to 140–150, and further rises to nearly 200 by Frame 430. At Frame 760, due to silkworm movement, the dense region shifts to the lower-left area, where the ID values increase to approximately 260.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. Example tracking results of WormSORT on MST-100.

(a) Frame 4. (b) Frame 200. (c) Frame 430. (d) Frame 760.

https://doi.org/10.1371/journal.pcbi.1014410.g008

Fig 9 shows that the number of active tracked targets remains close to the true population size throughout MST-100. whereas the cumulative number of unique IDs increased steadily from approximately 100 to about 240, and the maximum assigned ID exceeded 250 at the end of the sequence. This indicates that the long-sequence degradation of WormSORT is mainly caused by cumulative identity fragmentation and ID switches, rather than substantial target loss. Combined with the qualitative results in Fig 8, these failures are primarily associated with prolonged overlap in densely populated regions, and are further aggravated by abrupt head swings that are not well captured by the Kalman filter.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 9. Temporal evolution of tracking behavior of WormSORT on the MST-100 sequence.

https://doi.org/10.1371/journal.pcbi.1014410.g009

Notably, WormSORT and other compared methods exhibit relatively low IDF1 and AssA scores on MST-100, indicating that heavy spatial crowding significantly degrades tracking performance. In densely distributed regions, frequent positional overlap leads to high IoU between targets, while appearance features become insufficiently discriminative. This inevitably results in identity switches. These observations suggest that severe crowding is the primary factor limiting tracking accuracy. Therefore, when applying MOT to silkworm trajectory analysis, avoiding excessively dense scenarios is essential for achieving reliable tracking performance.

4.3 Ablation analysis

To further validate the contribution of each component in WormSORT, ablation experiments were conducted on the MST-50 and MST-100 datasets. Method (a) first performs IoU-based matching followed by appearance-feature matching. Based on this baseline, IoU-padding was introduced in Method (b), feature-padding in Method (c), and both strategies were combined in Method (d), which corresponds to the full WormSORT framework.

As shown in Table 3, the baseline method (a) already achieved strong performance on MST-50, with a MOTA of 96.54%, IDF1 of 92.97%, and HOTA of 87.01%, indicating that the IoU-first matching strategy is effective for silkworm MOT. This result is consistent with the motion characteristics of silkworms, whose trajectories are relatively smooth and do not exhibit abrupt displacement between adjacent frames. Under such conditions, IoU provides reliable association cues, whereas appearance features are less discriminative when targets are densely distributed or partially overlapping.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Ablation results of WormSORT.

https://doi.org/10.1371/journal.pcbi.1014410.t003

Introducing IoU-padding in Method (b) led to consistent improvements, particularly on the more challenging MST-100 dataset. Compared with Method (a), MOTA increased from 91.71% to 92.00%, IDF1 from 56.59% to 63.15%, HOTA from 55.66% to 63.05%, AssA from 40.91% to 48.14%, and DetA from 76.65% to 82.73%. On MST-50, the gains were smaller but still observable in IDF1, HOTA, and AssA. These results indicate that IoU-padding effectively preserves high-confidence unmatched trajectories and detections for subsequent association, thereby reducing missed matches and improving both association quality and detection reliability.

By contrast, feature-padding alone in Method (c) produced mixed results. On MST-50, IDF1 increased slightly from 92.97% to 93.68%, but MOTA and HOTA decreased to 95.03% and 86.35%, respectively. A similar trend was observed on MST-100, where Method (c) improved IDF1 and AssA relative to the baseline but remained inferior to Method (b) across all metrics. This suggests that although feature-padding can provide supplementary cues for identity preservation, its contribution is limited when appearance features are insufficiently discriminative, especially in crowded scenes where severe overlap occurs among silkworm targets.

When IoU-padding and feature-padding were jointly applied in Method (d), the best overall performance was achieved on both datasets. On MST-50, Method (d) yielded 96.37% MOTA, 93.65% IDF1, 91.33% HOTA, 89.85% AssA, and 92.93% DetA. On MST-100, the corresponding values reached 94.93%, 64.12%, 64.81%, 49.37%, and 85.22%, respectively, all of which were the best among the four methods. Notably, the combined strategy produced the largest gains in HOTA, AssA, and DetA, indicating that the two padding mechanisms are complementary: IoU-padding primarily strengthens motion- and location-based association, while feature-padding further improves identity continuity in ambiguous cases.

Overall, the ablation results demonstrate that IoU-first matching is well suited to silkworm tracking, and that the proposed padding mechanisms, particularly when combined, substantially improve tracking robustness and association accuracy. These findings confirm the effectiveness of the full WormSORT design.

Table 4 reports the impact of the IoU-padding threshold δ on tracking performance. Overall, δ = 5 provides the best trade-off across both datasets, achieving the highest IDF1, HOTA, and AssA on MST-100, while maintaining stable performance on MST-50. When δ is too small (δ = 3), the candidate set becomes overly restrictive, leading to missed associations; conversely, a larger threshold (δ = 7) introduces more ambiguous candidates, slightly degrading association quality. These results indicate that a moderate motion constraint is essential for balancing recall and reliability in IoU-based candidate selection.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Ablation results of different δ values.

https://doi.org/10.1371/journal.pcbi.1014410.t004

Table 5 presents the ablation results for different feature-padding thresholds ε. As ε increases, stricter cosine similarity constraints improve identity discrimination, with ε = 0.8 achieving the highest IDF1 on both datasets. However, this also leads to reduced MOTA, HOTA, and DetA, as fewer candidates are retained for matching. In contrast, lower thresholds (ε = 0.4, 0.6) allow more flexible matching and yield better overall detection–association coverage. These findings suggest that feature-padding primarily refines identity consistency, but requires a careful balance to avoid over-restricting valid associations.

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Ablation results of different ε values.

https://doi.org/10.1371/journal.pcbi.1014410.t005