idTracker: Unsupervised Animal Tracking

• idTracker is a legacy multi-animal tracking tool that operates fully unsupervised, relying on background subtraction and thresholding without manual annotations.
• It employs a segmentation-based detection method and overlap association, complemented by a GUI for ease of use in research applications.
• Benchmark evaluations highlight its limitations in dense, long-term tracking while emphasizing its practical advantage in low annotation scenarios.

idTracker is a traditional multi-animal tracking (MAT) system for tracking unmarked animals in video, characterized in recent benchmarking as a fully unsupervised animal-tracking methodology for both detection and identification and included as a representative legacy MAT baseline because it is widely used, user-friendly, and GUI-based (Bibinbe et al., 15 Sep 2025). In contemporary comparative studies, it occupies a distinctive position: it requires no supervised detector training and no keypoint or bounding-box supervision in its tracking pipeline, yet it is increasingly evaluated against supervised multi-object tracking (MOT) systems and newer prompt- or foundation-model-based trackers that target the same long-term identity-preservation problem under crowding, occlusion, and appearance ambiguity (Bibinbe et al., 15 Sep 2025).

1. Historical and methodological position

In recent livestock-tracking benchmarking, idTracker is explicitly placed in the category of unsupervised detection + unsupervised identification, with no keypoints or bounding boxes used as manual supervisory signals inside the tracking pipeline (Bibinbe et al., 15 Sep 2025). This categorization is central to its identity within the field: idTracker is treated not as a modern detector-centric MOT pipeline, but as an established MAT tool whose principal appeal lies in its accessibility and low annotation burden.

The software’s standing is reinforced by its continued inclusion alongside other long-used animal-tracking packages. The benchmark on pig tracking states, “We included two widely used MAT methods: DeepLabCut and idTracker,” and further notes that, for MAT, “idTracker (version 5.2.12) and ToxTrac are SOTA approaches widely used by researchers since they are user-friendly with graphical user interfaces to assist the user” (Bibinbe et al., 15 Sep 2025). In that framing, idTracker remains relevant not because it defines the current performance frontier, but because it is a common operational choice for animal researchers who prefer established software over more technical MOT stacks.

A second strand of the literature positions idTracker as the precursor to later “tracking by identification” systems. The idtracker.ai paper states that it extends the earlier idTracker idea of “tracking by identification” while replacing major components with deep learning and an adaptive multi-stage training strategy (Romero-Ferrero et al., 2018). This places idTracker within a lineage of identity-centric animal tracking in which long-term individual persistence, rather than only short-term frame-to-frame linkage, is the core design objective.

2. Core tracking mechanism

The technical characterization used in current benchmarking describes idTracker as a completely unsupervised methodology, even for detection (Bibinbe et al., 15 Sep 2025). The pipeline, as summarized there, is as follows:

1. each frame is normalized to its mean to remove the background;
2. image regions (“blocks”) with normalized intensity above or below a threshold are treated as candidate animals;
3. animals are detected by subtracting the background using an average model;
4. segmentation is performed using thresholding and a filter based on the given size of the animals;
5. animals are tracked using the overlap of blocks across frames;
6. a deep neural network learns animal features in case of occlusion (Bibinbe et al., 15 Sep 2025).

This description makes clear that idTracker does not depend on a supervised object detector trained on species-specific boxes or a supervised keypoint estimator. Its detection stage is segmentation-like rather than detector-like, and its association stage begins from block overlap rather than from a learned tracking-by-detection stack. A plausible implication is that idTracker is best understood as a background-subtraction and thresholding pipeline augmented with learned appearance features for difficult cases, especially occlusions.

At the same time, the benchmarking paper is explicit about what it does not provide for reproducibility. For idTracker, it reports the software version and the high-level algorithmic description, but it does not report exact threshold values, size-filter settings, the background-model strategy, the deep neural network architecture, occlusion-handling parameters, identity confidence thresholds, post-processing settings, or manual-correction protocol (Bibinbe et al., 15 Sep 2025). Thus, idTracker is technically identifiable at the level of workflow and design philosophy, but not fully parameterized in that study.

3. Data assumptions, supervision profile, and software usage

In the pig benchmark, idTracker was run directly on the target video without a train/validation/test split, precisely because it is unsupervised (Bibinbe et al., 15 Sep 2025). This distinguishes it from DeepLabCut, which used a 90% train / 10% test split, and from supervised MOT pipelines that relied on a YOLOX pig detector trained with an 80% train / 20% validation partition (Bibinbe et al., 15 Sep 2025). For idTracker, the operational assumption is therefore not curated labeled training data, but a video amenable to unsupervised foreground extraction and subsequent identity reasoning.

The same benchmark makes explicit that idTracker’s own inputs do not include supervised pig boxes or keypoints. Instead, its practical requirements are tied to the detection scheme itself: a background model, thresholding, and a size filter based on expected animal size are at least implicit prerequisites (Bibinbe et al., 15 Sep 2025). This suggests that successful use depends materially on the compatibility between the scene and a segmentation-driven detector.

The broader idTracker lineage, as represented by idtracker.ai, illuminates the kind of recording conditions under which identity-centric unmarked-animal tracking is expected to succeed. idtracker.ai assumes that the number of animals is known, that there exists at least one portion of the video where all animals are visible and separated, that frame rate is high enough for blobs from the same animal to overlap in consecutive frames, that image quality is sufficient for visual identification, and that grayscale segmentation is possible (Romero-Ferrero et al., 2018). It also works best with good contrast, homogeneous illumination, minimal shadows and reflections, enough pixels per animal, and enough movement to expose appearance variability (Romero-Ferrero et al., 2018). These assumptions are stated for idtracker.ai rather than idTracker itself, but they indicate the recording logic that underlies the broader “tracking by identification” family.

Operationally, idTracker’s strongest practical attribute in the pig benchmark is software usability rather than benchmark accuracy. The paper identifies a user-friendly, standalone graphical interface as a key advantage of traditional MAT tools like idTracker, especially for users without a technical background (Bibinbe et al., 15 Sep 2025).

4. Benchmark performance on long-term pig tracking

The most detailed recent empirical characterization of idTracker comes from the pig-tracking benchmark in “Multi-animal tracking in Transition: Comparative Insights into Established and Emerging Methods” (Bibinbe et al., 15 Sep 2025). The evaluation used a 10-minute, 25 FPS video of 15 active growing-finishing pigs in a pen, for a total of 15,000 frames with 782 annotated frames (Bibinbe et al., 15 Sep 2025). The dataset contained both keypoints and bounding boxes; each pig had two ears, neck, tail, and nose annotated when visually available, and bounding boxes were generated as minimum bounding rectangles around the keypoints, yielding 11,730 annotations of pigs including bounding boxes, five keypoints, and identity (Bibinbe et al., 15 Sep 2025).

Tracking evaluation used IDF1, MOTA, identification F1 score, identification recall, and identification accuracy. The paper defines IDF1 verbally as an identity-aware F1 score that measures how consistently identities are maintained over time, and MOTA as a global metric that penalizes missed detections, false positives, and identity switches (Bibinbe et al., 15 Sep 2025). For unsupervised methods, detection was separately evaluated with detection F1 score, detection recall, and detection accuracy using $IOU > 0$ (Bibinbe et al., 15 Sep 2025).

On this benchmark, idTracker produced the following reported values:

Evaluation group	Metric	Value
Tracking	IDF1	0.16
Tracking	MOTA	0.32
Tracking	Identification F1	0.10
Tracking	Identification recall	0.09
Tracking	Identification accuracy	0.10
Detection	Detection F1	0.86
Detection	Detection recall	0.77
Detection	Detection accuracy	0.97

These were the lowest overall tracking results among all evaluated methods in that study (Bibinbe et al., 15 Sep 2025). The contrast is sharp: ByteTrack achieved IDF1 0.79 and MOTA 0.73; DeepSORT achieved IDF1 0.74 and MOTA 0.72; Track-Anything and PromptTrack also exceeded idTracker by large margins on identity-related scores (Bibinbe et al., 15 Sep 2025). Even relative to other MAT tools, idTracker underperformed: DeepLabCut with five keypoints reached IDF1 0.52 and MOTA 0.84 (Bibinbe et al., 15 Sep 2025).

The benchmark’s qualitative evidence is consistent with the numerical results. A figure caption describes an idTracker snapshot showing several undetected animals, whereas Track-Anything and PromptTrack are shown with all animals detected (Bibinbe et al., 15 Sep 2025). The discussion states that “the main limitation of most unsupervised approaches lies in the detection stage,” and specifically that Track-Anything and PromptTrack outperform idTracker in both recall and accuracy, with superior detection capabilities improving tracking performance (Bibinbe et al., 15 Sep 2025). The paper therefore locates idTracker’s weakness not only in long-term identity maintenance, as reflected by IDF1 $=0.16$, but also earlier in the detection stage under dense pig-pen interactions.

5. Comparison with related systems and paradigms

The recent pig benchmark concludes that MOT approaches consistently outperform MAT methods in terms of tracking performance, even in long-term scenarios (Bibinbe et al., 15 Sep 2025). In that comparison, idTracker represents the traditional unsupervised MAT paradigm, while ByteTrack, DeepSORT, and Cross-Input Consistency represent supervised MOT pipelines using a supervised pig detector, and Track-Anything and PromptTrack represent newer methods that do not require detector retraining on the pig dataset (Bibinbe et al., 15 Sep 2025). The practical conclusion of that study is that traditional MAT tools, including idTracker, are no longer the strongest option for crowded long-duration pig tracking.

A different comparison emerges in the ToxTrac paper. There, idTracker is characterized as a system that extracts characteristic fingerprints from each animal and matches them to trajectories, but is also described as very computationally heavy, as relying on a non-general separation stage, and as not scaling well with the number of animals, video length, or video resolution (Rodriquez et al., 2017). Those claims are comparative positioning by the ToxTrac authors rather than a neutral benchmark consensus, but they are part of how idTracker has been situated in the animal-tracking software literature.

The ToxTrac paper also provides a direct benchmark in which average detection rates in non-occluded experiments were 99.2% for ToxTrac versus 95.9% for idTracker (Rodriquez et al., 2017). At the same time, it states that in occluded fish settings idTracker obtains better results because of a resegmentation stage that improves performance for experiments with multiple animals with fish-like shape (Rodriquez et al., 2017). This comparison highlights a recurring trade-off: idTracker is more identity-centric and, in some overlap-heavy settings, benefits from more specialized handling of animal merges, whereas other systems may be faster or more general-purpose.

The relationship to idtracker.ai is different again. idtracker.ai is not a synonym for idTracker but an extension of the earlier idTracker idea using two deep networks—a deep crossing detector (DCD) and an identification CNN (idCNN)—together with an adaptive cascade of self-training protocols (Romero-Ferrero et al., 2018). Under good recording conditions, idtracker.ai reports very high validated accuracy in large collectives, including zebrafish and flies, often above 99.9% and up to 100 individuals (Romero-Ferrero et al., 2018). In historical terms, this means that the idTracker lineage did not end with the original unsupervised segmentation-and-overlap formulation; it evolved into deeper identity-learning systems designed to handle much larger groups.

6. Limitations, misconceptions, and prospective directions

A common misconception is to treat idTracker as interchangeable with newer deep identity trackers or with generic MOT pipelines. The literature instead presents it as a distinct legacy MAT system: fully unsupervised in the benchmarked formulation, dependent on background subtraction, thresholding, size filtering, and overlap-based association, and valued primarily for accessibility and absence of supervised detector training (Bibinbe et al., 15 Sep 2025). idtracker.ai belongs to the same lineage but is a different system with a much more elaborate learned architecture and adaptive training strategy (Romero-Ferrero et al., 2018).

Another misconception is to infer that because idTracker is unsupervised, its deployment is parameter-free. The pig benchmark does not provide a full hyperparameter sheet, but its own algorithm summary implies dependence on thresholding, background modeling, and a provided animal-size filter (Bibinbe et al., 15 Sep 2025). The same study also leaves unreported whether GPU was used, whether manual intervention or correction was allowed, and whether the GUI was used interactively during the run (Bibinbe et al., 15 Sep 2025). For rigorous reproducibility, those omissions matter.

The most important substantive limitation identified in recent evaluation is that idTracker is not competitive for accurate long-term pig identity tracking in dense livestock footage. In the 10-minute, 15-pig benchmark, it had the lowest overall performance among all evaluated approaches, with the discussion attributing much of the failure to the detection stage and showing several missed animals qualitatively (Bibinbe et al., 15 Sep 2025). Its main remaining practical strength is therefore usability rather than state-of-the-art accuracy.

Recent work on robust point tracking suggests one possible direction for future identity-preserving animal tracking. DiTracker shows that pretrained video diffusion transformers can provide strong temporal correspondence through motion, blur, occlusion, and reappearance, especially on difficult reappearance-heavy scenarios (Son et al., 23 Dec 2025). Because point-level persistence is not the same as full instance identity tracking, this does not directly replace idTracker. A plausible implication, however, is that idTracker-like systems could benefit from stronger low-level temporal correspondence modules for resolving ambiguity among similar-looking individuals, particularly under repeated occlusion and reappearance (Son et al., 23 Dec 2025).