PanAf: Astronomy and Ecology Perspectives

• PanAf is a context-sensitive acronym that denotes both the Pan-African Asteroid Search Campaign in astronomy and the Pan African Programme in primatology.
• In astronomy, PanAf drives a continent-wide citizen science initiative for asteroid detection using FITS images from surveys like Pan-STARRS and the Catalina Sky Survey.
• In ecology and computer vision, PanAf supports wildlife behavior analysis with challenging camera-trap datasets designed to address class imbalance and out-of-distribution robustness.

Searching arXiv for papers using the term “PanAf” and closely related usages. PanAf is a field-dependent acronym rather than a single stable referent. In astronomy, it denotes the Pan-African Asteroid Search Campaign, a continent-wide citizen-science effort coordinated through the Pan-African Citizen Science e-Lab (PACS e-Lab) in partnership with the International Astronomical Search Collaboration (IASC). In ecology and computer vision, it denotes the Pan African Programme: The Cultured Chimpanzee and several derivative datasets built from its camera-trap archive, including PanAf-500, PanAf20k, PanAf500, and PanAf-FGBG (Marcel et al., 2024, Brookes et al., 2023, Brookes et al., 28 Feb 2025). The term therefore has to be interpreted from disciplinary context: in one literature it refers to planetary defense and continent-scale educational infrastructure, and in another it refers to field primatology data and machine-learning benchmarks for wildlife behavior analysis.

1. Terminological scope

Across recent arXiv usage, “PanAf” has at least two principal meanings and several adjacent near-matches. The astronomy usage is explicitly defined by the paper “Pan-African Asteroid Search Campaign: Africa’s Contribution to Planetary Defense” (Marcel et al., 2024). The ecology/computer-vision usage derives from the Pan African Programme: The Cultured Chimpanzee, which supplies camera-trap data for behavior recognition, background-bias analysis, and related wildlife-vision tasks (Suessle et al., 2023, Mueller et al., 15 Sep 2025).

Usage of “PanAf”	Domain	Representative source
Pan-African Asteroid Search Campaign	Astronomy, citizen science, planetary defense	(Marcel et al., 2024)
Pan African Programme: The Cultured Chimpanzee	Primatology, ecology, wildlife computer vision	(Brookes et al., 2023)
PanAf-derived datasets such as PanAf-FGBG	Wildlife behavior recognition and OOD analysis	(Brookes et al., 28 Feb 2025)

This dual usage is the central fact needed for disambiguation. A plausible implication is that references to “PanAf” without field markers are intrinsically ambiguous, especially when astronomy, ecology, and machine learning are discussed in the same corpus.

2. PanAf as the Pan-African Asteroid Search Campaign

In astronomy, PanAf is the Pan-African Asteroid Search Campaign, described as the continent-wide asteroid-search effort coordinated through PACS e-Lab in partnership with IASC. PACS e-Lab is the African organizing platform, while IASC supplies the astronomical data and the broader validation framework. The title frames the campaign as “Africa’s Contribution to Planetary Defense.” PACS e-Lab is described as having been established “to promote hands-on activities in astronomy & space science through citizen science and Soft Astronomy research in Africa,” and since December 4, 2020 it has served as IASC’s biggest partner in Africa (Marcel et al., 2024).

The institutional lineage is explicit. IASC was founded in 2006 by Dr. Patrick Miller at Hardin-Simmons University as an educational outreach program providing access to “high-quality astronomical datasets” for discovering and tracking asteroids, comets, and related objects. The collaboration between PACS e-Lab and IASC was formally facilitated through the African Astronomical Society, specifically by Dr. Charles Takalana. By the time of writing in July 2024, PACS e-Lab had spread the project to more than 40 African countries and organized the search every month, with academic-year calendars for 2023/2024 and 2024/2025.

The workflow is operationally specific. Each participating team has a minimum of two people. Teams receive datasets containing four FITS images captured at intervals of 30–60 minutes. The images come from Pan-STARRS and the Catalina Sky Survey, pass through IASC, and are analyzed with Astrometrica, a Windows-based program for asteroid detection and measurement. Participants load Minor Planet Center data, register Astrometrica using IASC-issued codes, load the four images, activate the “Blink” function, and generate a Minor Planet Center (MPC) report. IASC then evaluates whether the report is a valid asteroid candidate or a false detection. The staged pathway runs from monthly participation certificates, to expert evaluation about a week later, to preliminary discoveries, then to provisional discoveries over six months to one year, and eventually to cataloging and possible naming.

The campaign is both scientific and educational. The paper emphasizes that automated pipelines in surveys such as Pan-STARRS and CSS can miss some asteroids because of “fill factor effects, signal-to-noise requirements, and non-optimal sky-plane motion.” PanAf therefore enlarges manual inspection capacity within a recognized validation chain. Its headline scale indicators are 58 groups, 40 countries, 595 active asteroid citizen scientists, and 32 provisional discoveries. The abstract states that “About 595 citizen scientists from over 40 countries” had been engaged, while the body reports “52 enlisted discoveries” across African citizen scientists overall, of which “32 are affiliated with PACS e-Lab.” The campaign also reports that asteroid search is “the leading project with 90.7% participation” among PACS e-Lab activities, with East Africa leading, followed by North Africa, West Africa, Southern Africa, and Central Africa. For 2024/2025, IASC granted PACS e-Lab “35 slots,” meaning about 35 groups can participate in the research each month, and the stated ambition is to recruit “thousands of citizen scientists from all 54 African countries.”

3. PanAf as the Pan African Programme in field primatology

In ecology and wildlife computer vision, PanAf refers to the Pan African Programme: The Cultured Chimpanzee, a large camera-trap effort centered on wild great apes. One paper states that the full Pan-African dataset contains about 20,000 videos from 39 study sites across 15 African countries (Brookes et al., 2023). Another describes PanAf as having collected data at more than 40 temporary and collaborative research sites across Central and West Africa using motion- and infrared-equipped camera traps, producing over 600,000 one-minute video clips (Suessle et al., 2023).

The primary biological design was chimpanzee research rather than generalized wildlife re-identification. This matters because later methodological papers use PanAf footage for tasks that were not the original collection target. The leopard-identification study states explicitly that PanAf was originally designed to study chimpanzees, not leopards, and that the camera locations were chosen to suit chimpanzee behavior rather than species-specific leopard survey design (Suessle et al., 2023). This makes the archive a difficult robustness test bed rather than an optimized benchmark.

The observational conditions are consistently described as challenging. PanAf footage includes low nighttime illumination, frequent black-and-white infrared imagery, blur, low quality, partial body visibility, variable animal distance from camera, diverse poses, and fixed backgrounds. These conditions are central to why PanAf became useful in machine learning: the corpus supplies naturally difficult, in-the-wild data rather than laboratory or zoo recordings. A plausible implication is that performance claims on PanAf are best interpreted as robustness claims under field constraints rather than narrow benchmark saturation.

4. PanAf-derived datasets and benchmark design

Several benchmark datasets in wildlife computer vision inherit the PanAf name. PanAf-500 is described as a curated benchmark subset of the larger Pan-African dataset. In one formulation it consists of 500 videos, about 180,000 manually annotated frames, frame-by-frame annotations, full-body locations of great apes, and one of nine behavioural action classes (Brookes et al., 2023). The same paper emphasizes strong class imbalance, with class frequencies spanning roughly two orders of magnitude, and highlights average per-class accuracy (C-Avg) as particularly important.

A later behavior-recognition paper describes PanAf in a somewhat different benchmark configuration. There, PanAf is a dataset of camera-trap videos from 18 field sites in tropical Africa that capture chimpanzees and gorillas. It is divided into PanAf20k, containing 20,000 coarsely annotated videos, and PanAf500, containing 500 videos with fine-grained bounding box, track, and frame-wise behavior annotation. Each video is 15 seconds long. For the downstream task, PanAf500 is used under a snippet-wise single-label classification protocol: training and evaluation operate on 16-frame snippets, each showing a sustained single behavior of a single ape, with one predicted behavior label per snippet. The nine classes are sitting, walking, standing, hanging, climbing up, sitting on back, running, camera interaction, and climbing down, listed in descending order of frequency, again indicating class imbalance (Mueller et al., 15 Sep 2025).

PanAf-FGBG is a more specialized derivative benchmark designed to isolate the role of environmental context. It features 21 hours of footage, 5,070 video pairs, clips of 15 seconds, and footage from 389 individual camera locations, 14 national parks / research sites, and 6 African countries. Its defining property is paired sampling: each foreground video containing a chimpanzee is matched to a background video with no chimpanzee from the same camera location. The dataset provides two views: an overlapping camera-location view and a disjoint camera-location view. It also supplies multi-label behaviour annotations, unique camera ID, and detailed textual scene descriptions. The goal is to measure how much behavior-recognition systems exploit behavior-correlated backgrounds, and how that affects in-distribution versus out-of-distribution generalization (Brookes et al., 28 Feb 2025).

Taken together, these benchmarks show that “PanAf” in computer vision is not a single fixed dataset schema. It can refer to a broad camera-trap program, a nine-class action-recognition subset, a larger unlabeled-plus-labeled great-ape corpus, or a paired foreground/background benchmark for OOD analysis. This suggests that exact variant naming—PanAf-500, PanAf500, PanAf20k, or PanAf-FGBG—is methodologically significant.

5. Methods and empirical results developed on PanAf data

PanAf data have supported several distinct methodological lines. In wildlife re-identification, the paper “Automatic Individual Identification of Patterned Solitary Species Based on Unlabeled Video Data” uses a leopard subset spanning 2011–2018, with 210 videos from eight field sites and 68 unique camera locations. The pipeline extracts frames, uses MegaDetector for animal detection, relies on SIFT features via HotSpotter for matching, and clusters videos into putative individuals to avoid the open-set problem. On this PanAf subset it found 116 matches, of which 97 were correct, yielding a success rate of 83.6% (Suessle et al., 2023).

In great-ape behavior recognition, the paper “Triple-stream Deep Metric Learning of Great Ape Behavioural Actions” introduces a triple-stream architecture over RGB appearance, optical flow, and DensePose-C chimpanzee body-part information. On PanAf-500, the best top-1 result is 85.86% with triple-stream plus element-wise multiplication fusion. For long-tail performance, the best average per-class accuracy is 65.66% with weight balancing, compared with prior literature at 42.33% (Brookes et al., 2023). The paper frames this as the first deep metric-learning system for great-ape behavioural actions and explicitly treats embedding geometry as scientifically informative.

A later line of work uses self-supervised transfer. “Domain-Adaptive Pretraining Improves Primate Behavior Recognition” starts from a pretrained V-JEPA model and applies domain-adaptive pretraining (DAP) on unlabeled PanAf500 + PanAf20k. On PanAf500, the reported results are 83.68% Top-1 and 57.75% C-Avg for V-JEPA (no DAP), and 87.24% Top-1 and 56.37% C-Avg for V-JEPA (DAP). The paper therefore reports a +6.15 percentage point gain over the previous best published Top-1 result of 81.09% from MViTV2, and a +3.56 percentage point improvement over plain V-JEPA. At the same time, class-average accuracy does not improve with DAP on PanAf, which the paper treats as an open challenge connected to imbalance and tail classes (Mueller et al., 15 Sep 2025).

PanAf-FGBG turns the archive into a benchmark for shortcut learning. The paper reports that BG-only models trained on paired background videos remain highly predictive: across architectures, background-only performance is never below about 70% of FG-only uAP and never below about 65% of FG-only mAP. It further introduces a latent-space background compensation rule,

$$z^{\mathcal{F}-\mathcal{B}} = z^{\mathcal{F}} - (1-\alpha)\cdot z^{\mathcal{B}},$$

and reports OOD gains on the disjoint camera-location split of +5.42% mAP for 3D R50 and +3.75% mAP for MViT-V2 under linear scheduling (Brookes et al., 28 Feb 2025). This makes PanAf a test bed არა only for recognition accuracy but for quantifying background dependence and location-shift failure modes.

6. Disambiguation, misconceptions, and adjacent usages

A common misconception is that “PanAf” always denotes a single program or dataset. The arXiv record considered here does not support that reading. In astronomy it is the Pan-African Asteroid Search Campaign; in ecology and computer vision it is the Pan African Programme and its derivative benchmarks. These are unrelated enterprises that share an acronym but differ in domain, data modality, and scientific objective (Marcel et al., 2024, Suessle et al., 2023).

Another potential confusion arises from near-matches in medical imaging. The paper “Pan-FM: A Pan-Organ Foundation Model with Saliency-Guided Masking for Missing Robustness” introduces Pan-FM, not PanAf, even though it is a pan-organ foundation model and thus a lexical near neighbor (Wu et al., 8 May 2026). The PDAC detection paper on the PANORAMA dataset explicitly states that it does not mention PanAf explicitly (Deng et al., 20 Feb 2026). The fatty-pancreas ultrasound paper addresses non-alcoholic fatty pancreas disease (NAFPD), also called pancreatic steatosis or fatty pancreas, but its topic is not an entity named PanAf (Anghel et al., 8 May 2026). These papers are relevant only as disambiguation boundaries.

The most stable encyclopedic reading, therefore, is that PanAf is a context-sensitive acronym whose meaning is discipline-specific. In one literature it names an African-led asteroid-search infrastructure connected to planetary defense. In another it names a pan-African great-ape field program and a family of camera-trap benchmarks that have become important for wildlife re-identification, long-tail behavior recognition, self-supervised adaptation, and OOD robustness. This suggests that precise expansion—Pan-African Asteroid Search Campaign versus Pan African Programme: The Cultured Chimpanzee—is not optional but necessary for technical clarity.