Impact of Background on Classification Performance of Camera Trap Images

Traditionally classification is performed on originally collected images due to its computational simplicity. Preprocessing steps are usually conducted before classification. This research aims to examine what effects segmenting out the background as well as replacing the background with natural landscapes has on classification performance for out-of-the-box state-of-the-art models. This paper summarizes research in image classification and the impact of impala background on the performance of a zero-shot MobileNetv3 classification model utilizing ImageNet weights. Most of the research is experimental in nature and involves the utilization of DSAIL-Porini images and labels, an open-source wildlife image dataset. The methodology involved the input of impala images into a pretrained model for classification and assessing the model's performance. The input images were all single-class, impala images. Segment Anything Model (SAM) and You Only Look Once version 8 (YOLOv8) detection models were used to remove the image backgrounds with the impalas as the foreground and a similar background with undetectable objects was also used to access the model's performance on the third experiment. The results show that upon removal of the background, the model's accuracy dropped significantly from 18.15% to 0.11%. Upon swapping the background with a similar conservancy background, there was a less significant drop in performance, 2.05%. This implies a significant effect of the background on classification performance.