Choosing an Appropriate Platform and Workflow for Processing Camera Trap Data using Artificial Intelligence

Vélez, Juliana; Castiblanco-Camacho, Paula J.; Tabak, Michael A.; Chalmers, Carl; Fergus, Paul; Fieberg, John

doi:10.48550/arxiv.2202.02283

Cited by 9 publications

(14 citation statements)

References 21 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Such a check is crucial, as machine learning models usually decrease dramatically in accuracy when they are applied to new data (Schneider et al, 2020). Thus, automatically classifying images that are completely new to the model, for example images taken after the model has been trained, can result in poorer model performance and the verification of automatic image labels has been emphasized by several authors (Christin et al, 2021;Vélez et al, 2022). In our case study, the original model classified the new images with a prediction accuracy over 90 %, which is very high when a classification model is transferred to new data (Schneider et al, 2020).…”

Section: Discussionmentioning

confidence: 99%

“…Most of the avialable tools focus on either automatic or manual image classification, although a combination might often be most appropriate. In many cases, neural networks will rather accelerate manual classification than completely replace it (Vélez et al, 2022;Greenberg) because the transferability of neural networks to new images is known to be problematic (Norouzzadeh et al, 2018) and classification models usually perform better for some species than for others. Thus, the importance of a verifying automatic image labels has been emphasized (Christin et al, 2021).…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

A semi-automatic workflow to process camera trap images in R

Böhner

Kleiven

Ims

et al. 2022

Preprint

View full text Add to dashboard Cite

Camera traps have become popular for monitoring biodiversity and animal populations. Artificial intelligence is increasingly used to automatically classify large image data sets produced by camera traps and many tools that incorporate machine-learning models for automatic image classification have been developed over the last years. However, it is still challenging to combine tools for automatic classification with other tools for processing camera trap images and to adapt these tools to a specific study. Therefore, we propose a semi-automatic workflow for processing camera trap images in R. The workflow includes managing raw images, automatic image classification, a quality check of automatic image labels as well as the possibilities to retrain the model with new images and to manually review subsets of images to correct image labels. We illustrate the workflow with a case-study from the small mammal monitoring program of the Climate-ecological Observatory for Arctic Tundra. We first trained a classification model for small mammals and then transferred the model to new images, including images from newly established camera traps. We could show that retraining the original model with a small number of new images increased model performance and therefore highlight the importance of verifying automatic image labels when a model was transferred to new images. Furthermore, retraining the original model also decreased the time needed for manually reviewing images and correcting image labels substantially. Thus, the proposed workflow results in a data set with high accuracy and minimizes time needed for labeling images manually. This is especially useful for long-term monitoring where new images have to be processed continuously and methods have to be adapted over time. We provide all R scripts and the classification model for small mammals to make the workflow accessible to other ecologists.

show abstract

Section: Discussionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

A semi-automatic workflow to process camera trap images in R

Böhner

Kleiven

Ims

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Our models are effective at identifying and counting species of interest and at distinguishing empty images from those with animals. There are several software options available for automatically processing camera trap images using deep learning computer vision algorithms (Vélez et al 2022). Our R package is most similar in functionality to MegaDetector (Beery et al 2019), a powerful and effective algorithm that detects, classifies, and counts humans, non-human animals, and vehicles in camera trap images.…”

Section: Discussionmentioning

confidence: 99%

“…Currently, software options exist for applying object detection algorithms to camera trap images, specifically, Wildlife Insights (Wildlife Insights 2021), MegaDetector (Beery et al 2019), Conservation AI (Chalmers et al 2019), AnimalFinder (Price Tack et al 2016), and ClassifyMe (Falzon et al 2020). Each of these methods have strengths and limitations, which are discussed in detail in a recent review of these software packages (Vélez et al 2022). Despite the availability of detection algorithms for identifying animals in camera trap images an R based tool is currently not available that can be integrated into workflows.…”

Section: Introductionmentioning

confidence: 99%

CameraTrapDetectoR: Automatically detect, classify, and count animals in camera trap images using artificial intelligence

Tabak

Falbel²,

Hamzeh

et al. 2022

Preprint

Self Cite

View full text Add to dashboard Cite

Motion-activated wildlife cameras, or camera traps, are widely used in biological monitoring of wildlife. Studies using camera traps amass large numbers of images and analyzing these images can be a large burden that inhibits research progress. We trained deep learning computer vision models using data for 168 species that automatically detect, count, and classify common North American domestic and wild species in camera trap images. We provide our trained models in an R package, CameraTrapDetectoR. Three types of models are available: a taxonomic class model classifies objects as mammal (human and non-human) or avian; a taxonomic family model that recognizes 31 mammal, avian, and reptile families; a species model that recognizes 75 domestic and wild species including all North American wild cat species, bear species, and Canid species. Each model also includes a category for vehicles and empty images. The models performed well on both validation datasets and out-of-distribution testing datasets as mean average precision values ranged from 0.80 to 0.96. CameraTrapDetectoR provides predictions as an R object (a data frame) and flat file and provides the option to create plots of the original camera trap image with the predicted bounding box and label. There is also the option to apply models using a Shiny Application, with a point-and-click graphical user interface. This R package has the potential to facilitate application of deep learning models by biologists using camera traps.

show abstract

“…These models are used to help ecologists with labeling and detecting animals. Some researchers directly apply object detection networks, such as ResNet18 [ 10 ], GoogLeNet [ 11 ], Faster R-CNN [ 12 , 13 , 14 ], YOLO series [ 15 , 16 , 17 ], and AlexNet [ 18 ], for species recognition. To achieve better performance, a few researchers have tried to improve the networks or fused several networks.…”

Section: Introductionmentioning

confidence: 99%

Contour-Based Wild Animal Instance Segmentation Using a Few-Shot Detector

Tang

Zhao

et al. 2022

Animals

View full text Add to dashboard Cite

Camera traps are widely used in wildlife research, conservation, and management, and abundant images are acquired every day. Efficient real-time instance segmentation networks can help ecologists label and study wild animals. However, existing deep convolutional neural networks require a large number of annotations and labels, which makes them unsuitable for small datasets. In this paper, we propose a two-stage method for the instance segmentation of wildlife, including object detection and contour approximation. In the object detection stage, we use FSOD (few-shot object detection) to recognize animal species and detect the initial bounding boxes of animals. In the case of a small wildlife dataset, this method may improve the generalization ability of the wild animal species recognition and even identify new species that only have a small number of training samples. In the second stage, deep snake is used as the contour approximation model for the instance segmentation of wild mammals. The initial bounding boxes generated in the first stage are input to deep snake to approximate the contours of the animal bodies. The model fuses the advantages of detecting new species and real-time instance segmentation. The experimental results show that the proposed method is more suitable for wild animal instance segmentation, in comparison with pixel-wise segmentation methods. In particular, the proposed method shows a better performance when facing challenging images.

show abstract

Choosing an Appropriate Platform and Workflow for Processing Camera Trap Data using Artificial Intelligence

Cited by 9 publications

References 21 publications

A semi-automatic workflow to process camera trap images in R

A semi-automatic workflow to process camera trap images in R

CameraTrapDetectoR: Automatically detect, classify, and count animals in camera trap images using artificial intelligence

Contour-Based Wild Animal Instance Segmentation Using a Few-Shot Detector

Contact Info

Product

Resources

About