Source-linked AI summary
BossouChimpanzee: Long-term Chimpanzee Video Dataset
Daniel Schofield, Susana Carvalho, Vladimir Iashin, Andrew Zisserman, Max Bain, Arsha Nagrani, David Ng, Claudia Sousa, Boniface Zogbila, Jules Doré, Dora Biro, Misato Hayashi, Tetsuro Matsuzawa
TL;DR
Long-term naturalistic records provide a basis for studying wild chimpanzee behaviour and cognition. This paper describes the BossouChimpanzee archive and releases an annotated 74-hour subset; the archive spans 1988–2018 and the subset covers 23 identified individuals. The initial release is a resource for behavioural research and machine-learning work on audiovisual data from the wild.
Problem
Long-term naturalistic records of wild chimpanzees are valuable for examining behaviour, tool use, and the evolution of cognition in the wild.
Source: §1 · p. 2Method
Results
BossouChimpanzee70h comprises approximately 74 hours of video with audio, depicting 23 identified chimpanzees across six sampled years.
Source: §1 · p. 2Takeaways & Limitations
The archive and released subset offer resources for behavioural research and machine-learning work on audiovisual data from wild chimpanzees.
Source: §1 · p. 2, §4 · p. 8BossouChimpanzee70h is an initial subset focused on six years, leaving substantial portions of the archive, particularly earlier recordings, unexplored.
Source: §4 · p. 7
Abstract
from arXiv · showhide
We describe the BossouChimpanzee video dataset, a unique long-term visual record of wild chimpanzees at an outdoor laboratory for field experiments in Bossou, Guinea, spanning three decades (1988-2018) and comprising over 1,200 hours of continuous video recordings collected through collaborative fieldwork and research. In this paper, we outline the history and scientific contributions of the experimental paradigm and video archive, provide key statistics and details on the structure of the main video dataset, and release an initial ~74h snapshot, BossouChimpanzee70h, covering 23 identified individuals focused on chimpanzee individual and action recognition, ahead of the full video resource. This dataset represents a valuable resource for cognitive and behavioural research in ethology and a rich benchmark for training and evaluating machine learning models on audiovisual data from the wild.
1 Introduction
BossouChimpanzee is a decades-long audiovisual record of wild chimpanzees in a quasi-experimental outdoor laboratory, where stone-tool and nut availability could be controlled. The paper introduces the archive and releases an annotated 74-hour subset for individual and action recognition.
Overview
- The archive documents wild chimpanzees at Bossou from 1988 to 2018, capturing tool use in a naturalistic setting designed to study behavioural variation.Controlling stone-tool and nut availability supported investigation of how behaviour emerges under ecological conditions.Source: §1 · p. 2
Released subset
- BossouChimpanzee70h releases approximately 74 hours of audiovisual footage in 226 files and 7.8 million frames, depicting 23 identified chimpanzees.The recordings sample six years: 2000, 2004, 2006, 2008, 2012, and 2013.Source: §1 · p. 2
Annotations
- The subset has 7.3M face boxes grouped into 12.6k tracks, 2.3M body boxes across 3.6k tracks covering 13 individuals in 2012–2013, and nearly 11k action annotations across 11 individuals in 2012–2013.Of these, 8.4k action annotations are linked to body tracks.Source: §1 · p. 2
Paper scope
- The paper outlines the field site's history, data-collection protocols, dataset structure, research implications, and future directions for ethology and machine learning.The stated aim is a collaborative research ecosystem spanning AI, ethology, and conservation.Source: §1 · pp. 2–3
2 Dataset Context: The Outdoor Laboratory
Bossou's outdoor laboratory combined long-term observation with controlled provisioning of naturally available tools and nuts. Its archive spans decades, captures tool use and social behaviour, and varies in recording format and visual quality.
2.1 History and Scientific Context
- Bossou's long-term research documented tool use, social dynamics, and other behaviours, while the community faced isolation, habitat pressures, disease, and population decline.No females were recorded joining the community for more than forty years; by 2022, three chimpanzees remained.Source: §2.1 · p. 3
2.2 Field Experimental Paradigm
- Researchers provisioned naturally available stones and nuts at the outdoor laboratory to study nut-cracking under semi-structured conditions in a forest clearing.Annual experiments ran from 1986 to 2018, with some years unavailable for data collection.Source: §2.2 · p. 3
2.3 Recording Protocol
- Fixed-tripod cameras recorded sessions from chimpanzee arrival to departure, capturing tool choices alongside vocalising, play, aggression, and affiliative interactions.Depending on the year, recordings used single or multiple cameras with wide-angle and zoomed views.Source: §2.3 · pp. 3–4
2.4 Source Collection and Digitisation (1988–2018)
- The 1988–2018 archive contains 1,491 video files and over 1,200 hours, assembled from analogue tapes and later digital recordings.Older footage typically has 720 × 480 resolution, while later digital video is higher fidelity; some analogue tapes show degradation or signal loss.Source: §2.4 · p. 4
3 BossouChimpanzee70h
BossouChimpanzee70h is a six-year video subset of the wider archive, with annotations spanning chimpanzee identities, bodies, and nut-cracking actions. The section describes how these annotations were produced and organized.
3 BossouChimpanzee70h
- The ∼74-hour subset covers 226 videos from six years and provides face- and body-level detection, tracking, identity, frame-level identity, and audiovisual action annotations.The six years are 2000, 2004, 2006, 2008, 2012, and 2013.Source: §3 · p. 4
3.1 Face Annotations
- Face annotations use a trained detector and KLT tracking, with an expert assigning one identity per track across 19,512 tracks and 24 labels.The labels cover 23 individuals plus one faulty-track label; the release includes four years.Source: §3.1 · p. 5
3.2 Body Annotations
- The reprocessed body annotations comprise ∼3.6k tracks and 2.3M body instances across 13 individuals, with uncertain identity transfers retained as Unknown.SAM 3 detections are linked with ByteTrack, and identity labels are transferred by accumulated overlap with original tracks; the released action annotations remain linked to earlier body tracks.Source: §3.2 · p. 6
3.3 Frame Annotations
- Frame-level identity labels enumerate chimpanzees visible in a frame; 0.9M frames from 27 videos in 2012–2013 are released.Further frame labels could be mined from other videos with face- and body-box identity annotations; Bain et al. [26] used such labels to train counting and then body recognition without body tracks.Source: §3.3 · pp. 6–7
3.4 Audiovisual Spatio-Temporal Action Annotations
- Nut-cracking annotations cover on-screen spatio-temporal and off-screen audio-only events, labeling strikes, failures, cracking bouts, crack success, and subsequent eating.The iterative process uses body boxes to seed labels, an audio-only ResNet-18 [43] pretrained on VGGSound [44] to propose strikes, and expert verification; 10,895 action annotations include off-screen strikes.Source: §3.4 · p. 7
4 Future work and BossouChimpanzee
BossouChimpanzee70h is an initial subset focused on six years, leaving substantial archive material, especially earlier recordings, unexplored. The authors identify further annotation and research opportunities and plan to release the full archive after content filtering.
4 Future work and BossouChimpanzee
- The initial subset covers six years on which most existing annotations focus, leaving substantial archive portions, particularly earlier recordings, unexplored.The full archive is planned for release once required content filtering is complete.Source: §4 · p. 7
- Further annotation and benchmarking opportunities span individual identification, pose estimation, action and interaction recognition, and multimodal audiovisual analysis.Proposed research questions include long-term behavioural change, ontogeny, learning, cultural transmission, life histories, ageing, sociality, and tool use.Source: §4 · pp. 7–8
A.1 BossouChimpanzee Video Archive
The archive appendix describes media distributions and clarifies how its per-year counts differ from those for the BossouChimpanzee70h subset.
A.1 BossouChimpanzee Video Archive
- Archive media summaries include videos and duration per year, video frame-rate distribution, and audio-channel distribution.The table assigns recordings to the field season identified by the first video date; seasons spanning a new year are filed under the starting year.Source: §A.1 · p. 11, Table A1 · p. 11
- Table A1 counts digitised video files, each corresponding to one archived tape; BossouChimpanzee70h instead counts excerpted clips.Subset per-year counts can therefore exceed archive counts while its total duration remains a strict subset.Source: Table A1 · p. 11
A.2 BossouChimpanzee70h
The appendix identifies media-attribute distributions for BossouChimpanzee70h and a video-duration histogram whose vertical axis is count.
A.2 BossouChimpanzee70h
- The video-duration histogram uses count as its vertical-axis label.The supplied passages give no histogram values or further axis details.Source: §A.2 · p. 12
- The appendix names video duration as a plotted distribution for BossouChimpanzee70h.No duration values are stated in the supplied passages.Source: §A.2 · p. 12, Table A2 · p. 12
- Table A2 reports media-attribute distributions for BossouChimpanzee70h.The supplied passage does not specify the table's individual attributes or values.Source: Table A2 · p. 12