Eugênio Dias Ribeiro Neto

Thesis

AI-based detection and tracking of individual animal interactions using camera-trap.

PhD in Computer science at the University of Montpellier (January 2024 - December 2026).

Keywords: Animal re-identification, Camera traps, Dog re-identification, Deep learning, Domain generalization, Wildlife monitoring.

Director: Marc Chaumont - Associate professor at the University of Southern Brittany and Researcher at the IRISA and at the LIRMM.

Co-director Michel De Garine-Wichatitsky - Researcher at the CIRAD UMR ASTRE research unity.

Co-supervisor: Gérard Subsol - CNRS Researcher at the LIRMM.

Co-supervisor: Hélène Guis - Researcher at the CIRAD UMR ASTRE research unity.

Photo of Eugênio Dias presenting the paper 'An improved architecture for part-based animal re-identification through semantic segmentation distillation' at the WACV 2026 in Tucson, USA in march 2026.
Our paper 'Background-invariant re-identification of dogs from camera-trap videos in non-controlled environments' was published at the journal Ecological Informatics in december 2025. The image shows the first page of the paper.
Our paper 'An improved architecture for part-based animal re-identification through semantic segmentation distillation' was accepted at the IEEE/CVF WACV 2026 in november 2025. The image shows the first page of the paper.
With my colleagues G. Picaud and G. Fourret (both on the left of this photo), we were placed 2nd out of 250 participants in the PINKCC international challenge of ovarian cancer segmentation challenge. The photo shows us in the prize ceremony.
I presented my research at the MIPS research center week of the university of Montpellier. The image shows on top an example of dog detection results and below an example of clustering.
I spent one month in Thailand in Cambodia to participate in the data collection of the SEAdogSEA project. The photo was tajen in Cambodia, it contains the data collection team composed by members from the CIRAD-ASTRE team and members of the Institut Pasteur du Cambodge.
I presented my research at the MIPS research center week of the university of Montpellier. The image shows on top an example of dog detection results and below an example of clustering.

I am currently in my 3rd (and last) year of PhD at the I2S doctoral school of the University of Montpellier. I am based at the ICAR team of the LIRMM laboratory. I am also affiliated to the CIRAD UMR ASTRE research unity.

My thesis is part of the SEAdogSEA, an interdisciplinary project between french and asian research institutions. The goal of the project is to study the contact patterns between dogs and their habitat to better comprehend their behavior and role as zoonotic disease transmitters. To this end, camera-traps were installed across three Asian countries, more specifically, Indonesia, Thailand and Cambodia. The thesis focuses on the development of an AI-based pipeline to automate the analysis of the large volumes of data generated by the camera-traps, with a particular focus on re-identification.

Please check the resume of the thesis below and our publications.

Resume of the thesis

The study of interactions between animals in their natural environment is essential for understanding disease transmission dynamics and for designing effective conservation and epidemiological control strategies. In recent years, camera trap networks have produced large volumes of video footage, but analyzing interactions between individuals of a given species requires processing massive amounts of data, which remains a major bottleneck.

This thesis presents an automated video-processing pipeline based on artificial intelligence algorithms, organized around three main stages: detection of individuals of a given species, multi-object tracking within a video sequence, and re-identification (Re-ID), that is, recognizing the same individual across different sequences captured either by the same camera trap at a different time or by another camera trap in the network.

Our work focuses on Re-ID in the wild conditions (less studied in the literature) characterized by large variations in viewpoint and illumination, occlusions, cluttered backgrounds, and, more generally, the substantial appearance changes that occur when camera traps are redeployed at the same location during acquisition campaigns separated by several months.

Our approach is developed and validated on free-roaming dogs filmed in Southeast Asia, a setting of high epidemiological relevance, while being designed to extend naturally to quadrupeds in general. Our first contribution addresses the lack of suitable datasets. We release two resources to the community: YT-BB-Dog, the largest dog Re-ID dataset to date (more than 2,700 individuals automatically extracted from YouTube videos), and SEAdogSEA-Indonesia, a multi-camera dataset spanning three years, comprising 183 dogs recorded by 47 camera traps over four acquisition sessions in an Indonesian village, enriched with geographic and socio-ecological metadata validated by both specialists and the dogs' owners.

Based on these datasets, we propose two new methods for animal Re-ID. BIFOR (Background-Invariant Feature extractOR) is a three-stage training procedure based on a novel mini-batch sampling strategy, designed to mitigate the domain shift that arises in domain-generalizable Re-ID, where a model trained on a source dataset must perform well on an unseen target domain. We then introduce PAW-ViT (Part-AWare animal re-identification Vision Transformer), which replaces the standard class token of a ViT with K body-part tokens. Each additional token specializes in a specific anatomical region through knowledge distillation from semantic segmentation, within a multi-task learning framework. The part tokens are then fused by an additional aggregation token into a single descriptor that, by construction, weights each part according to its visibility in the image and its importance for Re-ID. Both methods outperform state-of-the-art baselines, particularly in uncontrolled cross-camera scenarios. PAW-ViT further confirms the broader applicability of our framework to quadrupeds, achieving state-of-the-art results on datasets such as ATRW (Amur tigers) and YakReID-103 (yaks).

Finally, we benchmark our two methods together with a wide range of state-of-the-art algorithms on SEAdogSEA-Indonesia, in order to evaluate Re-ID performance in a realistic open-set scenario (no prior information is available about which individuals may reappear across campaigns) and over the long term (with individuals and environments evolving across years). This study covers the entire pipeline, from raw camera-trap footage to contact networks, providing a realistic framework intended for use by ecologists and epidemiologists.

Additional contributions