Evaluation of deep learning architectures for wildlife object detection: A comparative study of ResNet and Inception

Wildlife object detection plays a vital role in biodiversity conservation, ecological monitoring, and habitat protection. However, this task is often challenged by environmental variability, visual similarities among species, and intra-class diversity. This study investigates the effectiveness of two individual deep learning architectures ResNet-101 and Inception v3 for wildlife object detection under such complex conditions. The models were trained and evaluated on a wildlife image dataset using a standardized preprocessing approach, which included resizing images to a maximum dimension of 800 pixels, converting them to RGB format, and transforming them into PyTorch tensors. A ratio of 70:30 training and validation split was used for model development. The ResNet-101 model achieved a classification accuracy of 94% and a mean Average Precision (mAP) of 0.91, showing strong performance in extracting deep hierarchical features. The Inception v3 model performed slightly better, attaining a classification accuracy of 95% and a mAP of 0.92, attributed to its efficient multi-scale feature extraction through parallel convolutions. Despite the strong results, both models exhibited challenges when detecting species with similar visual characteristics or those captured under poor lighting and occlusion. Nonetheless, the findings confirm that both ResNet-101 and Inception v3 are effective models for wildlife object detection tasks and provide a reliable foundation for conservation-focused computer vision applications.

💡 Research Summary

This paper presents a systematic comparative study of two widely used convolutional neural network (CNN) architectures—ResNet‑101 and Inception v3—for wildlife object detection, a task that underpins modern biodiversity monitoring and conservation efforts. The authors assembled a custom wildlife image dataset comprising thousands of photographs captured across diverse habitats (savannas, forests, semi‑arid regions) and featuring a broad range of species, lighting conditions, poses, and occlusions. All images were uniformly pre‑processed: the longest side was resized to a maximum of 800 pixels, color channels were converted from BGR to RGB, and the data were transformed into PyTorch tensors. A 70 % training / 30 % validation split was employed to ensure a fair comparison between the two models.

Both networks were initialized with ImageNet‑pretrained weights and fine‑tuned using the same training configuration: cross‑entropy loss, Adam optimizer with a learning rate of 0.01, batch size of 32, and a maximum of 50 epochs with early stopping to avoid over‑fitting. Training was performed on an NVIDIA RTX 3090 GPU, taking roughly six hours for ResNet‑101 and five hours for Inception v3.

The evaluation protocol included overall classification accuracy, mean Average Precision (mAP), per‑class precision, recall, F1‑score, and confusion matrices. ResNet‑101 achieved 94 % accuracy and a mAP of 0.91, while Inception v3 slightly outperformed it with 95 % accuracy and a mAP of 0.92. Detailed class‑wise analysis revealed that both models performed well on large, well‑lit animals but struggled with small or partially hidden subjects, low‑light images, and species that share similar color patterns (e.g., certain antelopes vs. zebras). The confusion matrices highlighted frequent misclassifications in these challenging scenarios.

The discussion links the observed performance differences to architectural characteristics. ResNet‑101’s deep residual blocks facilitate the learning of high‑level hierarchical features, which benefits detection in cluttered scenes with prominent objects. Inception v3’s parallel convolutional paths and factorized kernels enable multi‑scale feature extraction, giving it a modest edge in detecting objects of varying sizes and in more complex visual contexts. Both models, however, share limitations related to class imbalance, insufficient illumination, and occlusion, suggesting that additional strategies—such as advanced data augmentation (color jitter, random cropping, MixUp), class‑weighting, or ensemble methods—could further improve robustness.

The paper situates its contributions within the broader literature, noting that prior works using ResNet‑50, VGG‑16, or Faster R‑CNN reported accuracies in the 90–93 % range and mAP values between 0.85 and 0.89. By moving to deeper (ResNet‑101) and more sophisticated (Inception v3) architectures, the authors achieve a 1–2 % absolute gain in both metrics, confirming the value of architectural depth and multi‑scale design for wildlife detection tasks.

In conclusion, the study demonstrates that both ResNet‑101 and Inception v3 are viable backbones for automated wildlife monitoring systems, delivering high accuracy and strong detection performance under standardized conditions. Nonetheless, real‑world deployments will need to address the identified challenges—poor lighting, vegetation occlusion, and visually similar species—through richer datasets, balanced class distributions, and possibly hybrid or lightweight models for on‑edge inference. The work provides a clear benchmark and methodological blueprint for future AI‑driven conservation projects.