Source-linked AI summary
Towards Automatic Wild Animal Monitoring: Identification of Animal Species in Camera-trap Images using Very Deep Convolutional Neural Networks
Alexander Gomez, Augusto Salazar, Francisco Vargas
TL;DR
Camera trapping generates thousands or millions of images, creating a manual analysis burden for wildlife researchers. This paper uses very deep convolutional neural networks across Snapshot Serengeti dataset variants to identify species under camera-trap conditions, reaching 88.9% Top-1 and 98.1% Top-5 accuracy. The results support automated recognition while exposing behavioral, annotation, and fine-grained classification limitations.
Problem
Camera-trap species identification remains unsolved because wildlife images are numerous and affected by challenging environmental and behavioral conditions.
Method
The paper evaluates six very deep convolutional neural networks on four Snapshot Serengeti dataset versions representing different camera-trap scenarios.
Results
88.9% Top-1 and 98.1% Top-5 accuracy were achieved on the evaluation set using a residual network topology.
Takeaways & Limitations
The results show that species recognition in camera-trap images can be automated and that the models are robust to wrong citizen annotations.
Takeaways & Limitations
Animal behavior remains uncontrolled, and annotation corruption raises unresolved questions about required dataset size and effects across training and evaluation.
Abstract
from arXiv · showhide
Non intrusive monitoring of animals in the wild is possible using camera trapping framework, which uses cameras triggered by sensors to take a burst of images of animals in their habitat. However camera trapping framework produces a high volume of data (in the order on thousands or millions of images), which must be analyzed by a human expert. In this work, a method for animal species identification in the wild using very deep convolutional neural networks is presented. Multiple versions of the Snapshot Serengeti dataset were used in order to probe the ability of the method to cope with different challenges that camera-trap images demand. The method reached 88.9% of accuracy in Top-1 and 98.1% in Top-5 in the evaluation set using a residual network topology. Also, the results show that the proposed method outperforms previous approximations and proves that recognition in camera-trap images can be automated.
1 Introduction
Camera trapping supports important wildlife research but produces thousands or millions of images that experts must analyze manually. The paper proposes automatic species identification using very deep convolutional neural networks on Snapshot Serengeti dataset variants.
- Camera traps support questions about species presence, behavior, abundance, distributions, predation, and population monitoring.
- Thousands or millions of camera-trap photographs are currently analyzed manually because automatic species identification is not used.An automatic system could accelerate professionals’ work and leave them more time for data analysis.
- Automatic species classification remains unsolved because camera-trap images contain challenging conditions.
- Very deep convolutional neural networks are evaluated on unbalanced, balanced, conditioned, and segmented Snapshot Serengeti datasets.These variants probe four issues: unbalanced samples, empty frames, incomplete animal images, and distant objects.
2 Related Work
Prior camera-trap recognition studies used handcrafted features or a shallow ConvNet on private datasets. The proposed approach increases learning capacity and evaluates multiple dataset conditions, including crops containing only animal parts.
- Yu et al. recognized 18 species with 82% accuracy using sparse coding spatial pyramid matching on a 7,196-image dataset.Their pipeline used dense SIFT, local binary patterns, sparse coding, and max pooling.
- Chen et al. classified 20 species with 38.31% accuracy using a six-layer ConvNet and automatic animal segmentation.
- The proposed approach evaluates unbalanced, balanced, conditioned, and segmented datasets with very deep AlexNet, VGGNet, GoogLeNet, and ResNet models.
- Unlike Yu et al.’s whole-body crops, the proposed manually segmented images can contain only some animal body parts, increasing classification complexity.
3 Towards animal monitoring in the wild
Camera-trap recognition must handle environmental degradation, unpredictable animal behavior, hardware effects, partial views, and fine-grained species differences. The paper addresses these conditions with deep ConvNets that transform image features from low- to high-level representations.
- 3 Towards animal monitoring in the wild: Camera-trap challenges comprise environmental conditions, animal behavior, and hardware limitations.
- 3.1 Challenges in camera trapping: Environmental changes can cause occlusion, poor illumination, overexposure, rain, and lens drops during long-term deployment.
- 3.1 Challenges in camera trapping: Unpredictable poses, partial bodies, and multiple species in one image can hide discriminative features and require localizing all animals.
- 3.1 Challenges in camera trapping: Blurred and overexposed animal images depend on camera hardware and selected parameters.
- 3.1 Challenges in camera trapping: Similar species and partial or unexpected poses make fine-grained classification harder.
- 3.2 Convolutional Neural Networks: Very deep ConvNets are used because their learning capacity and robustness support recognition from partial information.
- 3.2 Convolutional Neural Networks: ConvNets transform low-level features such as edges and orientation into high-level representations such as fur details and discriminative patterns.
- 3.2 Convolutional Neural Networks: AlexNet, VGGNet, GoogLeNet, and ResNet provide increasing-depth architectures with features including small filters, inception modules, or shortcut connections.
4 Experimental Framework
The experiments use multiple Snapshot Serengeti and Panama dataset configurations to test very deep ConvNets under class imbalance, foreground placement, segmentation, and cross-dataset conditions. Models are compared using validation accuracy, Top-1 and Top-5 accuracy, across 32 architecture–dataset experiments plus a Panama comparison.
- Datasets: Snapshot Serengeti was published in 2015 from 225 camera traps in Tanzania, with more than one million picture sets annotated through Consensus and Gold Standard subsets.The Consensus set was validated against the Gold Standard at 96.6% species-identification accuracy and was used to train and validate the models.
- Datasets: The Panama dataset contains 20 North American species, with 14,346 training images and 9,530 testing images produced using automatic segmentation.It was included to compare the best Snapshot Serengeti model with a previous approach.
- Datasets: The study selected 26 Snapshot Serengeti classes and constructed dataset versions to examine unbalanced data, balanced sampling, foreground placement, and manual segmentation.D1 contains all selected images; D2 samples 1000 training and 240 evaluation images per class; D3 uses foreground animals; D4 uses manually segmented images including partial animals.
- Models and experiments: Six very deep ConvNet architectures, including AlexNet, VGGNet, GoogLeNet, and ResNets, were evaluated to examine how depth affects camera-trap classification.One experiment was conducted for each architecture on each dataset, producing 32 experiments; fine-tuning was limited by hardware constraints.
- Evaluation and implementation: Validation accuracy was the performance metric, with Top-1 and Top-5 accuracy used to assess species ranking and classification assistance as the number of species increases.Images were resized for each topology, and final fully connected layers were replaced to classify 26 classes instead of the 1000 ImageNet classes.
- Evaluation and implementation: All architectures were ImageNet-pretrained, and fine-tuning used stochastic-gradient-descent back-propagation with reduced learning rates and step sizes to limit overfitting.The implementation used Caffe, and models, configuration files, and dataset versions were made publicly available for benchmarking.
5 Results
Across four Snapshot Serengeti dataset versions, recognition improved with deeper architectures and with cleaner or better-balanced data. The best result used D4 with architecture E, while fine-tuning improved performance but deeper fine-tuned models could not be tested.
- Dataset comparisons: D1 performed worst because even the deepest architectures could not handle its highly unbalanced data.Deeper models nevertheless outperformed less-deep models in this setting.
- Dataset comparisons: D2 produced higher accuracy than D1, but performance remained under 70% because images included empty frames and distant animals.The classifier may handle these cases only when they occur in small numbers.
- Dataset comparisons: 82.9% Top-1 and 96.8% Top-5 accuracy were reached on D3 after empty frames were removed.The results were much better than those for D1 and D2, and the deepest architectures again performed best.
- Dataset comparisons: 88.9% Top-1 and 98.1% Top-5 accuracy were achieved on D4 using architecture E.These were the best reported results across the dataset versions.
- Model comparisons: Fine-tuned architectures A and C outperformed black-box networks across all dataset versions.The authors suggest deeper fine-tuned architectures might perform better, but hardware limitations prevented testing that hypothesis.
- Model comparisons: Architecture E achieved high per-class accuracy on D4, with lower performance concentrated in fine-grained classification cases.Table 4 reports species-recognition performance by class on the D4-trained model.
6 Discussion
The discussion analyzes recurring misclassification patterns caused by insufficient visual information, fine-grained similarities, appearance variation, camera damage, and annotation errors. It also identifies data and imaging conditions needed to improve robustness.
- Discussion: The ConvNet may lack enough visual information to predict the correct species, specializing in particular animal parts whose presence influences classification.Misclassified examples include baboons as impalas and buffalo as wildebeest or hyenas.
- Discussion: Fine-grained errors occur because features such as a black body line are shared across Grant’s Gazelle, Thomson’s Gazelle, and impala.The black line is not a deterministic distinction between Grant’s and Thomson’s Gazelle, while similar back patterns further complicate Grant’s Gazelle–impala classification.
- Discussion: Top-5 accuracy remains high when Top-1 prediction is wrong, indicating that the correct species is often near the predicted answer.The authors interpret this as evidence that the ConvNet learned discriminative species patterns despite difficult images.
- Discussion: Age-related appearance variation can cause a buffalo cub to lack the distinctive pattern learned for adult buffalo.The discussion therefore calls for samples from cubs, females, and males of the same species.
- Discussion: Damaged cameras and grayscale imagery can bias predictions toward hyena or hippopotamus when animal shapes resemble those classes.The authors propose providing the hyena class with more color images.
- Discussion: Citizen-annotated training data can support robustness to annotation corruption when the training set has sufficient samples, but the required dataset size and corruption impact remain open questions.The discussion reports robustness on another citizen-annotated dataset while identifying further data-quality questions.
7 Conclusions
The conclusions present very deep convolutional networks as a promising approach for automatic species recognition in camera-trap images. Experiments across Snapshot Serengeti variants achieved high accuracy, while future work targets temporal information, animal-part training, and improved segmentation.
- 7 Conclusions: The method uses very deep convolutional neural networks and evaluates six architectures across four versions of the Snapshot Serengeti dataset.The dataset versions represent possible automated camera-trap scenarios.
- 7 Conclusions: The experiments show robustness to wrong citizen annotations and expose classification difficulties from age- or sex-related intra-class appearance variation.The authors state that highly accurate recognition is possible with sufficient data and an accurate segmentation algorithm.
- 7 Conclusions: Future experiments will evaluate bursts of camera-trap images as single instances using temporal information rather than analyzing each image independently.The authors expect sequence evaluation could address several previously described image conditions.
- 7 Conclusions: Future models will include animal parts in training images to support fine-grained classification and enable use of ImageNet images to expand training data.This proposal is motivated by the fine-grained nature of camera-trap recognition.
- 7 Conclusions: A short-term goal is to develop a segmentation algorithm that approximates manual segmentation.