The technical heart of the challenge lies in what is called re-identification. Object detection systems can draw boxes around every turkey in a single video frame, but behavior analysis demands something more: the ability to follow the same animal across frames, through occlusions and dense crowding, under a consistent identity. Multi-object tracking algorithms such as DeepSORT accomplish this by combining motion predictions with an appearance embedding—a numerical fingerprint, typically 128 numbers long, that captures what an individual looks like. The trouble is that the appearance model shipped by default with DeepSORT, a network called mars-small128 with roughly 2.8 million parameters, was trained on pedestrians. Humans are easy to tell apart by clothing, texture, and body structure; turkeys in a barn, uniform in color and shape, offer almost none of these cues. The result is that the generic model loses much of its discriminative power the moment it is pointed at poultry.

Sen and Lutz asked whether a compact, domain-specific embedding, trained on actual turkey data, could recover the discrimination that generic models lose while remaining light enough for on-farm deployment. Their answer is a Siamese convolutional neural network of just 714,000 parameters—about a quarter the size of the pedestrian baseline—that maps a cropped image of a turkey to a 128-dimensional vector on the unit hypersphere. The architecture was not designed in one stroke but arrived at through a careful, ablation-driven refinement process, in which each design decision had to justify itself through measurable gains in retrieval accuracy and training stability on held-out, unseen identities.

Three design elements survived that process. First, the network’s residual blocks use pre-activation ordering, a arrangement in which batch normalization and activation functions precede each convolution, keeping the skip connection free of non-linear transformations and improving gradient flow during training. Second, every residual block is augmented with squeeze-and-excitation channel attention, a small gating module that learns to emphasize the most informative feature channels—in this case, apparently, the subtle plumage coloration and body markings that do distinguish one turkey from another. Third, and most impactful, the channel count of the network’s deepest stage was doubled from 64 to 128, placing representational capacity where it matters most rather than in the final projection layer. An embedding head built around a batch-normalization neck, a technique borrowed from person re-identification that decouples metric and classification feature spaces, rounds out the design.

Training combined two complementary objectives: a batch-hard triplet loss, which pulls images of the same bird together in embedding space while pushing images of different birds apart by mining the hardest positive and negative pairs within each batch, and an auxiliary identity-classification loss with label smoothing, which regularizes the network on the modest training set of 74 turkey identities. The underlying dataset, extracted from overhead video of working commercial barns, comprised 106 unique turkey identities and 709 image crops, split so that the validation and test animals were never seen during training—a realistic open-set protocol, since any deployed system will inevitably encounter birds it has never learned from.

The results are striking. On an exhaustive pairwise evaluation of unseen test identities, the domain-specific embedding outperformed mars-small128 on every discrimination metric. The area under the receiver operating characteristic curve rose from 0.890 to 0.943, accuracy from 0.849 to 0.911, and recall by more than ten percent. Most telling was the separation between the similarity distributions of same-bird and different-bird pairs: the gap between their means grew by 73.5 percent, and a separability index improved by roughly half. Because false identity associations in tracking stem directly from overlap between these distributions, this cleaner separation is precisely the property that should translate into more stable identities downstream. Bootstrap confidence intervals confirmed that the gains were not an artifact of the small test pool. A t-SNE visualization of the embedding space showed tight, well-separated clusters for most individuals, with residual overlap confined to the genuinely most similar birds—an honest reflection of the problem’s intrinsic difficulty.

Just as important is what the small model does not cost. With 3.9 times fewer parameters and a footprint of 2.80 MiB versus 10.72 MiB, the custom network runs slightly faster per embedding than the baseline while using comparable GPU memory. Interpretability analyses using Grad-CAM showed that the network concentrates its attention on the bird’s body—the head–neck junction and upper back in standing birds, the breast and flank in resting ones—rather than on barn litter or background clutter, consistent with the squeeze-and-excitation design steering the model toward plumage and body-structure cues.

When the embedding was plugged into a DeepSORT tracking pipeline and tested on three commercial barn sequences, identity-related metrics improved consistently: mean IDF1, a strict measure of identity preservation, rose from 81.3 to 83.2 percent, and association accuracy improved, while detection accuracy remained unchanged—confirming the gain came from better association, not better detection. A controlled stress test that randomly dropped 10 to 30 percent of detections revealed a subtlety: the proposed embedding prefers a more permissive association threshold than the baseline, and once each model was evaluated at its own optimal setting, the domain-specific model won in five of six tested conditions. The authors are candid that the improvement in these particular sequences is moderate, largely because the footage contains few prolonged occlusions—the very regime where a strong appearance model matters most.

They are equally candid about the study’s limits. Identity fragmentation remains severe in absolute terms: even the improved tracker used about 90 track identities to cover roughly 57 real animals, meaning a farm relying on raw track counts would still over-count its flock. The authors identify this fragmentation, along with the need for larger-scale validation across more barns, flocks, and lighting conditions, as the primary remaining obstacle to reliable individual-level monitoring. Still, the central message stands and carries well beyond poultry: appearance representations learned for humans do not transfer to animals without loss, and the remedy is not a larger generic backbone but a smaller, domain-adapted one. For precision-livestock applications where compute and power are constrained, better identity discrimination and lower cost, it turns out, need not be in tension.

Subject of Research: A lightweight domain-specific appearance embedding for individual turkey re-identification in dense commercial barn environments using computer vision and multi-object tracking.

Article Title: A lightweight domain-specific appearance embedding for Turkey re-identification in dense barn environments

Article References: Sen, D., & Lutz, T. (2026). A lightweight domain-specific appearance embedding for Turkey re-identification in dense barn environments. Smart Agricultural Technology, 15, Article 102549. https://doi.org/10.1016/j.atech.2026.102549

Image Credits: AI Generated

DOI: 10.1016/j.atech.2026.102549

Keywords: turkey re-identification, precision livestock farming, computer vision, multi-object tracking, DeepSORT, Siamese network, appearance embedding, deep metric learning, triplet loss, squeeze-and-excitation attention, animal welfare monitoring, smart agriculture

The pilot study was conducted during the summer of 2023 at two very different UK institutions: the small Fenn Bell Conservation Project in Kent, where 76 visitors observed a pair of Asian small-clawed otters (Aonyx cinereus), and the much larger Twycross Zoo in the Midlands, where 155 visitors studied two Rothschild giraffe (Giraffa camelopardalis rothschildi). Participants were recruited through posters with QR codes at the enclosures and asked to watch the animals for two minutes before completing an online questionnaire built on a pre-defined ethogram, a structured list of species-typical behaviours including positive, negative and neutral welfare indicators. The questionnaire also invited optional sections on how the animals seemed to be feeling, what animal welfare meant to the visitors themselves, and how they had found the citizen science experience.

Technically, the design was deliberately cautious. Because visitor observations could not be paired with simultaneous video observations from the same individual, and because the two data sources differed in duration, the researchers did not treat the comparison as a validation exercise. Instead, they used systematically coded CCTV footage as a reference dataset, applying instantaneous scan sampling with one-minute intervals between 11am and 3pm, and summarising both datasets as hourly and daily presence or absence records. At the daily level, both data sources captured walking or running in the otters on all study days and chewing the cud, eating and drinking, standing and locomotion in the giraffe, while the largest discrepancies appeared for infrequent or harder-to-identify behaviours such as grooming and social interactions.

The psychological results proved equally instructive. Visitors scored the animals on eight qualitative terms, including content, relaxed, agitated, frustrated, comfortable, wary, confident and distressed, and then gave an overall welfare rating from 1 to 10. Spearman’s rank correlations revealed a coherent pattern at both sites: welfare scores rose alongside positively valenced adjectives such as content, comfortable, confident and lively or engaged, and fell alongside negative terms such as frustrated, agitated and wary. For the otters, the correlation with contentment was particularly strong, while for the giraffe even wary and agitated scores tracked downward with welfare ratings. Although participants were told in advance which terms indicated positive or negative welfare, the authors note these exploratory associations provide preliminary evidence of construct validity in public assessments of animal affective states.

Perhaps the most encouraging finding concerned how visitors conceptualised welfare itself. Asked to define or describe animal welfare, most respondents produced answers that mapped cleanly onto the Five Domains model, the contemporary framework covering nutrition, environment, health, behaviour and mental state. At Fenn Bell, 86 percent referenced physiological welfare, 64 percent mental state and 41 percent behaviour; at Twycross the figures were 93, 54 and 24 percent respectively. Fifteen otter visitors and thirteen giraffe visitors touched on all three broad themes, offering definitions such as meeting an animal’s emotional, physical, mental and health needs as naturally as possible through diet, environment and enrichment. The researchers observed, however, that visitors leaned towards input-based measures such as food and space rather than output-based measures such as behaviour, and that behavioural domains were the hardest for the public to articulate.

Where visitors learned their welfare concepts also carried a clear message for the zoo sector. Zoos themselves ranked among the most frequently cited sources of knowledge, with 45 of the Twycross respondents attributing their understanding to that zoo alone, ahead of television documentaries and educational facilities. The authors interpret this as evidence that educational material provided by zoos is being positively received, and they recommend that zoos place greater emphasis on positive behavioural interactions within their welfare science communication. Given that negative public perceptions of welfare can erode trust in zoos and reduce willingness to engage in conservation efforts, effective communication of welfare science is not a luxury but a strategic necessity for modern institutions.

Engagement levels were striking. At Twycross, where a researcher was present at the enclosure to encourage participation, an average of 27 visitors per day completed the survey; at the smaller Fenn Bell site, where participation was self-directed, the mean was 3.4 per day. Across both sites, 93 to 95 percent of respondents said they would take part in a similar activity again, and 79 percent overall described the process as rewarding. Crucially, the majority of participants were infrequent zoo visitors, with the largest group attending less than once a year, suggesting that citizen science could reach beyond the committed membership base of season-pass holders. Most participants also expressed confidence in their ability to identify the behaviours on the ethogram, which the researchers attribute to the deliberately simple design of the observation task.

The study is candid about the limitations of visitor-generated behavioural data. The otters were out of sight of the cameras for 67 percent of video observations, and both visitors and cameras faced blind spots and varying viewing opportunities. Behaviours that were infrequent or complex, such as positive social interactions in the otters or grooming in the giraffe, showed the greatest divergence between the two datasets. The authors stress that their analyses are exploratory descriptions of behavioural occurrence rather than measures of observer accuracy or reliability, and they call for future studies using time-matched designs, cameras positioned to replicate the visitor’s view, and controlled comparisons with trained observers to enable formal inter-rater reliability assessment. Training, structured prompts and simplified measures such as active versus inactive states are proposed as ways to improve data quality.

Beyond the data, the researchers argue, the process itself may deliver benefits that matter to the mission of the modern zoo. Watching animals closely is known to foster empathy and a sense of connectedness, visiting zoos has been linked to improved human wellbeing, and citizen science participation has been associated with gains in scientific understanding. Longer dwell times at exhibits may not automatically translate into learning, but they can deepen the visitor experience and support lasting connections with nature. As biodiversity losses reach unprecedented levels and conservation science becomes an ever more urgent discipline, the authors conclude that engaging zoo visitors in behavioural citizen science deserves far wider consideration, both as a source of data on how animals spend their opening hours and as a pathway to a more empathetic, conservation-minded public.

The study sits within a broader rethinking of what zoos are for. Traditional frameworks described zoos through four pillars of conservation, education, research and recreation, but more recent models have proposed adding a fifth pillar of wellbeing that applies to both animals and people, alongside a ‘sphere of influence’ recognising that zoos operate at local, national and global scales. Within that framing, communicating welfare science to the public is not peripheral but central, since public trust depends heavily on whether visitors believe the animals they are watching are thriving.

Citizen science itself is usually understood as having two distinct strands. In one, the priority is collecting usable data at scale, as in large wildlife monitoring programmes; in the other, the priority is engaging citizens, with data quality a secondary concern. The zoo project deliberately straddled both, and the tension between them explains much of the authors’ caution about the behavioural records. Previous zoo-based citizen science in the UK, such as campaigns encouraging visitors to log native species seen on site, has built ecological awareness but stopped short of fostering connections with the zoo’s own animals, a gap this study sought to address.

The choice of study species was pragmatic rather than incidental. Asian small-clawed otters and Rothschild giraffe were selected because they were consistently visible from public viewing areas, attracted longer dwell times, and could accommodate camera installations allowing simultaneous remote observation. Behavioural research is already the dominant form of zoo science, valued for capturing real-time information about animal experiences at low cost, yet large-scale data collection can be impractical for busy animal care staff. Distributing observation across millions of annual visitors offers one possible solution, provided questions of validity and reliability are resolved.

Earlier work also suggests the timing matters: visitors report more positive experiences and stronger feelings of connection after watching animals, particularly active ones, meaning the observation task itself may amplify the very engagement the projects aim to cultivate.

Subject of Research: Citizen science engagement of zoo visitors in animal behaviour and welfare assessment

Article Title: Exploring citizen science as a tool for engaging zoo visitors with animal welfare

Article References: Stephens, D., Wright, D., Rowden, L., Frost, N., Carter, A., Ward, S. J., & Williams, E. (2026). Exploring citizen science as a tool for engaging zoo visitors with animal welfare. Discover Animals, 3(1), Article 85. https://doi.org/10.1007/s44338-026-00247-2

Image Credits: AI Generated

DOI: 10.1007/s44338-026-00247-2

Keywords: citizen science, zoos, animal welfare, animal behaviour, Five Domains model, ethogram, visitor engagement, giraffe, Asian small-clawed otter, conservation education, public engagement, affective state