GreenCrossingAI: A Camera Trap/Computer Vision Pipeline for Environmental Science Research Groups

Andrea Osbon; Bernie Boscoe; Chandler Campbell; Karen Mager; Shawn Johnson

REVIEW 3 major objections 4 minor 17 references

GreenCrossingAI: A Camera Trap/Computer Vision Pipeline for Environmental Science Research Groups

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A small research group can run the pretrained MegaDetector on a local Windows GPU workstation to filter roughly one million camera-trap images, focusing human review on the ~4% that contain animals.

desk verdict A useful practice template for small labs running MegaDetector locally, but the central time-savings claim is unmeasured and the honest framing in the intro is contradicted by the unsupported conclusion. read the letter →

arxiv 2507.09410 v2 pith:4OJIF62G submitted 2025-07-12 cs.CV cs.LG

Bernie Boscoe , Shawn Johnson , Andrea Osbon , Chandler Campbell , Karen Mager This is my paper · ORCID

classification cs.CVcs.LG

keywords cameratrapsMegaDetectorwildlifemonitoringobjectdetectioncomputervisionpipelineon-premisemachinelearningTimelapseenvironmentalscienceworkflows

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

GreenCrossingAI is a documented attempt to show that a small environmental-science group, with student labor, a few thousand dollars, and no cloud budget, can run an existing computer-vision detector on its own computer to process camera-trap data. The authors report a working pipeline in which a pretrained detector (MegaDetector) runs on a local Windows workstation with a consumer GPU, and students review the bounding-box results inside their existing Timelapse labeling workflow. The pipeline handled roughly one million images from a wildlife-crossing study along Interstate 5, of which the authors say about 4% contain animals, and cut the manual review burden enough that a full workday can process about ten SD cards. The paper explicitly does not evaluate the model's accuracy on this dataset; its contribution is the workflow design and the demonstration that the approach is feasible under the group's constraints. A reader should care because this is a rare reported middle-ground solution between slow manual review and cloud AI services, and it describes the practical friction points in making ML usable for non-specialist ecology groups.

What carries the argument

The object that carries the argument is MegaDetector, a pretrained YOLOv5-based object detector that labels 'animal,' 'person,' and 'vehicle' and draws bounding boxes in camera-trap images. In this pipeline it acts as a filter rather than a final classifier: it produces detection JSON and images whose boxes let students see, at a glance, which thumbnails deserve a human look, while empty images are skipped. The second load-bearing component is Timelapse, the Windows-only labeling program the group already trusted, whose grid display and ten-minute event grouping are preserved so the new pipeline does not require retraining the researchers. Around these two sit the enabling hardware choices—a consumer Windows workstation with a 24 GB GPU, an SD card reader, and overnight backup to institutional storage—which make local inference (about 20 minutes per camera-location folder) fast enough for student workflows.

What would settle it

Count missed events: take a random sample of the human-verified animal images from the I-5 set, group them into the paper's ten-minute events, run MegaDetector over those images, and measure how many events contain no detection at all. If that event-level miss rate is not near zero, students would still have to scan blank images and the reported efficiency gain would not survive. A complementary timing test is to run the same end-to-end workflow with MegaDetector disabled and compare total student review time per SD card.

Watch

Extended reading notes

Core claim

The paper's central claim is that a low-resource, on-premise pipeline can bring state-of-the-art camera-trap detection to a small research group without cloud dependency or fine-tuning. In the reported implementation, MegaDetector runs in a Jupyter notebook on a Windows workstation and outputs JSON plus images with bounding boxes; students then open those results in the group's existing labeling software and review flagged images in a thumbnail grid, following the same ten-minute event rule they used when clicking through every image by hand. Over roughly four years of fieldwork the group amassed about 6.7 TB and approximately one million photos and videos from 20–35 cameras, of which only about 40,000 images contain animals; with the pipeline they recorded 8,295 unique events and 11,624 individual animals. The authors do not present a new model or a measured accuracy on their own data; they present an implementation path that other groups with similar budgets and skill levels could follow, and they argue that the remaining errors (false boxes on rocks or branches, missed frames in a burst) are tolerable because students can spot and correct them in the existing review interface.

Load-bearing premise

The time savings rest on the pretrained MegaDetector weights catching the animals in this particular camera-trap dataset well enough that students need to look only at flagged images; the paper does not measure that detection accuracy on the researchers' own data.

Editorial extensions

If this is right

Small ecology groups with a few thousand dollars and no cloud contract can adopt ML-assisted image review instead of clicking through every camera-trap frame.
Student effort shifts from finding animals to verifying detections and doing analysis, since roughly 96% of images are skipped.
The long-running I-5 dataset becomes a pre-construction baseline for the approved wildlife overpass, letting the group track animal activity before, during, and after construction.
Latency for ingesting a field session drops sharply: a 32 GB SD card that took 4–5 hours on an older laptop can be transferred, processed, and reviewed within a workday.
The documented failure modes (false boxes on rocks, missed frames in a sequence) are acceptable because at least one frame per event is usually detected, keeping event-level misses low.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Beyond the paper: the efficiency claim is conditional on event-level recall being high, which the authors never measure; a formal precision/recall audit against their existing human labels would turn the anecdotal 10,000-image check into a quantitative guarantee.
Beyond the paper: the same architecture—a pretrained detector on a local GPU plus an existing labeling GUI plus overnight backup—should transfer to other small wildlife groups, since the bottlenecks described (budget, Windows habits, student turnover, privacy of endangered-species locations) are widespread.
Beyond the paper: the power-measurement device already installed on the GPU means the promised solar-grid integration can be evaluated with data; publishing the pipeline's energy cost per image would let other groups compare local versus cloud processing on environmental grounds.
Beyond the paper: the cropping and zero-shot classification results the authors cite point to a natural upgrade path—adding a species classifier on top of MegaDetector's boxes without retraining the detector—which would further reduce manual labeling if the group ever wants it.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague — this is a practice report, not a research paper, and the most useful way to read it is as a template for a small lab that wants to run MegaDetector locally on Windows hardware. The authors describe a real deployment: an Alienware with an RTX 4090, a Jupyter notebook that wraps MegaDetector, Timelapse for labeling, and Box for storage. The requirements-gathering section (interviews with students, constraints around Windows/Box, reluctance to use Linux) is genuinely useful and not something you often see written down. Credit where due: the paper is honest about its own limits — it says in Section 1 that it does not evaluate model accuracy — and the workflow details (folder structures, upload times, event classification rules) are concrete enough that another group could replicate the setup.

The soft spot is the central claim. The conclusion says the pipeline 'significantly reduced the time and effort required for image labeling,' but there is no measurement to back that up. No baseline for how long manual labeling took, no count of missed animals from the 10,000-image check, no comparison against the 8,295 known events. The stress-test note has this right. The authors assert that 'the percentage of missed animals remains very low' without presenting the number. For a paper whose whole value is as a template, that's a real gap: another group reading this can't tell whether the pipeline actually works on this dataset or whether it just happened to feel faster. The paper would be stronger if the authors reported recall/precision on the 10,000 images they checked, or at minimum compared labeling time before and after.

Also minor: no repository link is given for the notebooks, which limits reproducibility. The hardware cost and specs are useful, but the claim that it runs on commodity hardware should be qualified — $2,750 plus a QNAP is not nothing.

Who is this for? People in small ecology or environmental science groups who are considering MegaDetector and want a step-by-step example. It is not for methodologists or ML researchers. I'd send it to a serious referee if I were the editor, because the template is useful and the deployment is genuine, but the referee should insist on either measurements or a rewritten conclusion that frames the efficiency gain as anecdotal rather than demonstrated. As is, it's a plausible practice writeup with an unsupported headline claim.

Referee Report

3 major / 4 minor

Summary. The paper describes GreenCrossingAI, an on-premise camera trap image processing pipeline centered on an Alienware workstation with an RTX 4090, MegaDetector, and Timelapse, developed for a small environmental science group at Southern Oregon University. It covers requirements gathering, hardware and software design, the workflow from field SD cards to analysis-ready CSV datasets, and implementation experience on roughly one million camera trap images. The paper explicitly states in Section 1 that it does not evaluate the accuracy of the model on the researchers' data, and its central claim in Section 11 is that integrating MegaDetector 'significantly reduced the time and effort required for image labeling.'

Significance. If the efficiency claim were substantiated with quantitative measurements, this paper would be a valuable template for small research groups seeking to adopt ML-assisted camera trap processing on local hardware, filling a clearly described gap between CPU-only and cloud-based workflows. The paper's strengths are its concrete description of a genuinely deployed system—including hardware specifications, workflow details, integration with Timelapse and Box, and practical maintenance considerations—and its explicit acknowledgment of limitations. However, the central claim of reduced labeling time is currently supported only by anecdote, and the paper declines to quantify detection accuracy, which is the premise on which the efficiency gain rests. As a result, the contribution is a useful practical field report rather than a demonstrated result.

major comments (3)

[§11 and §7.3] The conclusion claims that integrating MegaDetector 'significantly reduced the time and effort required for image labeling,' but Section 7.3 describes only the new workflow without any time measurements for labeling before versus after integration, and Table 1 reports SD card transfer times rather than labeling times. Without a quantitative baseline comparison of manual labeling versus MegaDetector-assisted labeling, the paper's central efficiency claim is unsupported.
[§8 and §1] The accuracy check of 10,000 images is reported only anecdotally: no numbers are given for recall, precision, missed-event counts, or per-species agreement relative to the 8,295 events and 11,624 individuals previously identified by human review, and Section 1 explicitly states that model accuracy on the researchers' data is not evaluated. The assertion that 'the percentage of missed animals remains very low' is therefore not backed by data, and the time-saving premise—that reviewers can rely on bounding boxes to skip blank images—remains unverified.
[§8] The paper mentions that MegaDetector produces false positives from swaying branches, rocks, or trees that resemble antlers, and that students correct these in Timelapse, but it does not quantify the frequency of such false positives or the additional review time they impose. Since the net labor savings depend on the balance between skipped blank images and corrected false positives, this omission further weakens the efficiency claim.

minor comments (4)

[§7.2] The text says 'A typical session runs as shown in Table ??' — this broken table reference should be replaced with 'Table 1'.
[§10] The paragraph beginning 'These insights are valuable for refining the use of MegaDetector...' is repeated verbatim twice in the Future Work section; one copy should be removed.
[References] Reference [3] contains a typo in the author name ('S, tefan Istrate' should be 'Ștefan Istrate'), and the paper does not state the MegaDetector version used; including the version would improve reproducibility.
[§6] The paper mentions that Jupyter notebooks are stored in GitHub, but no link or DOI is provided; adding a link would make the pipeline more reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports an implementation of an external detector and makes no fitted prediction that reduces to its inputs.

full rationale

The paper makes no derived quantitative prediction. It describes the integration of MegaDetector, an externally trained object detection model, into an existing manual camera-trap workflow. Section 1 explicitly disclaims accuracy evaluation on the researcher's own data: 'here we do not evaluate the accuracy of the model on the researcher's data, instead we focus on the implementation strategies, and meeting the computing needs of the researchers.' Section 8 reports only an anecdotal comparison of 10,000 images against human labels, with no fitted parameters, no recall/precision figures, and no model trained on the SOU dataset. The counts of events and individual animals cited in Section 8 come from prior manual labeling, not from the pipeline's detections. The central efficiency claim in Section 11 is therefore an implementation description rather than a prediction derived from the paper's own inputs. The lack of a quantitative accuracy evaluation and the absence of a measured manual-labeling baseline are evidentiary limitations and correctness risks, but they are not circularity. Self-references in the paper are project webpages and updates, not load-bearing mathematical premises. No equation, fitted parameter, or imported uniqueness theorem forces the stated outcome. Verdict: no significant circularity.

Assumptions & free parameters 0 free parameters · 2 assumptions · 0 invented entities

The central claim relies on external tooling (MegaDetector) that is trusted without on-site validation, and on a manual event-definition rule. No free parameters are fitted and no new entities are postulated.

assumptions (2)

domain assumption MegaDetector's pretrained weights detect animals in the SOU I-5 camera trap images well enough that human reviewers can rely on its bounding boxes.
The paper explicitly does not evaluate detection accuracy on this dataset (Section 1), yet the pipeline's time savings depend on MegaDetector catching animals and the review step only correcting occasional errors (Section 8).
domain assumption The ten-minute window rule defines independent animal events.
Used in Section 7.3 to convert images into events (8,295 events); this rule is inherited from the prior manual workflow and affects all downstream event and animal counts.

how reviews work

0 comments

Cite this review

Pith. "Pith review of GreenCrossingAI: A Camera Trap/Computer Vision Pipeline for Environmental Science Research Groups." pith.science (2026). https://pith.science/paper/4OJIF62G

@misc{pith2026250709410,
  author       = {Pith},
  title        = {Pith review of: GreenCrossingAI: A Camera Trap/Computer Vision Pipeline for Environmental Science Research Groups},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4OJIF62G}},
  note         = {Machine review of arXiv:2507.09410}
}

read the original abstract

Camera traps have long been used by wildlife researchers to monitor and study animal behavior, population dynamics, habitat use, and species diversity in a non-invasive and efficient manner. While data collection from the field has increased with new tools and capabilities, methods to develop, process, and manage the data, especially the adoption of ML/AI tools, remain challenging. These challenges include the sheer volume of data generated, the need for accurate labeling and annotation, variability in environmental conditions affecting data quality, and the integration of ML/AI tools into existing workflows that often require domain-specific customization and computational resources. This paper provides a guide to a low-resource pipeline to process camera trap data on-premise, incorporating ML/AI capabilities tailored for small research groups with limited resources and computational expertise. By focusing on practical solutions, the pipeline offers accessible approaches for data transmission, inference, and evaluation, enabling researchers to discover meaningful insights from their ever-increasing camera trap datasets.

Figures

Figures reproduced from arXiv: 2507.09410 by the authors.

**Figure 1.** A black-tailed deer (Odocoileus hemionus columbianus detected with GreenCrossingAI 2 The Role of Camera Traps and MegaDetector for Conservation Camera traps detect movement and capture images or videos in response. They function in various conditions, including infrared imaging at night or standard imaging during daylight hours, and often record additional metadata such as temperature, time, and other environmental … view at source ↗

**Figure 2.** Wildlife detected using GreenCrossingAI. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Flowchart depicting workflow from fieldwork to analysis-ready datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Timelapse user interface displaying detections. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 12 canonical work pages

[1]

Sara Beery, Dan Morris, and Siyu Yang. 2019. Efficient Pipeline for Camera Trap Image Review. arXiv:1907.06772

arXiv 2019
[2]

ElmorLabs. 2024. ElmorLabs PMD-USB (Power Measurement Device with USB). https://web.archive.org/web/20250123171559/https://elmorlabs.com/ product/elmorlabs-pmd-usb-power-measurement-device-with-usb/

arXiv 2024
[3]

Tomer Gadot, S, tefan Istrate, Hyungwon Kim, Dan Morris, Sara Beery, Tanya Birch, and Jorge Ahumada. 2024. To crop or not to crop: Comparing whole-image and cropped classification on a large dataset of camera trap images.IET Computer Vision 18, 8 (2024), 1193–1208. https://doi.org/10.1049/cvi2.12318

doi:10.1049/cvi2.12318 2024
[4]

Green Crossing AI. 2024. Green Crossing AI. https://greencrossingai.github.io/

work page 2024
[5]

Saul Greenberg. 2024. Home - Timelapse. https://timelapse.ucalgary.ca/

work page 2024
[6]

Emily Heller. 2024. SOU Students Support a Wildlife Project on I-5. https: //sou.edu/academics/sou-students-support-wildlife-project-on-i-5/

work page 2024
[7]

Lavista Ferres

Andres Hernandez, Zhongqi Miao, Luisa Vargas, Rahul Dodhia, Pablo Arbelaez, and Juan M. Lavista Ferres. 2024. Pytorch-Wildlife: A Collaborative Deep Learn- ing Framework for Conservation. https://doi.org/10.48550/arXiv.2405.12930

arXiv doi:10.48550/arxiv.2405.12930 2024
[8]

Shawn Johnson. 2024. GreenCrossingAI Project Update | WILDLABS. https: //wildlabs.net/discussion/greencrossingai-project-update-0

work page 2024

Show all 17 references

[9]

Labeled Information Library of Alexandria: Biology and Conservation. 2024. Data Sets. https://lila.science/datasets/

2024
[10]

Dan Morris. 2024. Everything I know about ML and camera traps. https: //agentmorris.github.io/camera-trap-ml-survey/

2024
[11]

Oregon Department of Transportation. 2024. Oregon Department of Transporta- tion : Project-Details : Projects : State of Oregon. https://www.oregon.gov/odot/ projects/pages/project-details.aspx?project=23100

2024
[12]

Oregon Wildlife Foundation. 2024. Southern Oregon Wildlife Crossing Coalition (SOWCC). https://myowf.org/sowcc

2024
[13]

Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. 2016. You Only Look Once: Unified, Real-Time Object Detection. https://doi.org/10.48550/arXiv. 1506.02640

2016 doi arXiv
[14]

Robin Steenweg, Mark Hebblewhite, Roland Kays, Jorge Ahumada, Jason T Fisher, Cole Burton, Susan E Townsend, Chris Carbone, J Marcus Rowcliffe, Jesse Whittington, Jedediah Brodie, J Andrew Royle, Adam Switalski, Anthony P Clevenger, Nicole Heim, and Lindsey N Rich. 2017. Scali...

2017 doi
[15]

Jiří Vyskočil and Lukas Picek. 2024. Towards Zero-Shot Camera Trap Image Categorization. https://doi.org/10.48550/arXiv.2410.12769

2024 doi arXiv
[16]

Castiblanco-Camacho, Michael A

Juliana Vélez, Paula J. Castiblanco-Camacho, Michael A. Tabak, Carl Chalmers, Paul Fergus, and John Fieberg. 2022. Choosing an Appropriate Platform and Workflow for Processing Camera Trap Data using Artificial Intelligence. https: //doi.org/10.48550/arXiv.2202.02283

2022 doi arXiv
[17]

Gosia Wozniacka. 2024. Oregon gets $33M to construct wildlife cross- ing over I-5 in southwest Oregon. The Oregonian (23 December 2024). https://oregonlive.com/environment/2024/12/oregon-gets-33m-to- construct-wildlife-crossing-over-i-5-in-southwest-oregon.html

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Sara Beery, Dan Morris, and Siyu Yang. 2019. Efficient Pipeline for Camera Trap Image Review. arXiv:1907.06772

arXiv 2019

[2] [2]

ElmorLabs. 2024. ElmorLabs PMD-USB (Power Measurement Device with USB). https://web.archive.org/web/20250123171559/https://elmorlabs.com/ product/elmorlabs-pmd-usb-power-measurement-device-with-usb/

arXiv 2024

[3] [3]

Tomer Gadot, S, tefan Istrate, Hyungwon Kim, Dan Morris, Sara Beery, Tanya Birch, and Jorge Ahumada. 2024. To crop or not to crop: Comparing whole-image and cropped classification on a large dataset of camera trap images.IET Computer Vision 18, 8 (2024), 1193–1208. https://doi.org/10.1049/cvi2.12318

doi:10.1049/cvi2.12318 2024

[4] [4]

Green Crossing AI. 2024. Green Crossing AI. https://greencrossingai.github.io/

work page 2024

[5] [5]

Saul Greenberg. 2024. Home - Timelapse. https://timelapse.ucalgary.ca/

work page 2024

[6] [6]

Emily Heller. 2024. SOU Students Support a Wildlife Project on I-5. https: //sou.edu/academics/sou-students-support-wildlife-project-on-i-5/

work page 2024

[7] [7]

Lavista Ferres

Andres Hernandez, Zhongqi Miao, Luisa Vargas, Rahul Dodhia, Pablo Arbelaez, and Juan M. Lavista Ferres. 2024. Pytorch-Wildlife: A Collaborative Deep Learn- ing Framework for Conservation. https://doi.org/10.48550/arXiv.2405.12930

arXiv doi:10.48550/arxiv.2405.12930 2024

[8] [8]

Shawn Johnson. 2024. GreenCrossingAI Project Update | WILDLABS. https: //wildlabs.net/discussion/greencrossingai-project-update-0

work page 2024