BugSR: Improving Tiny Instance Segmentation on the MassID45 Dataset

Automated biodiversity monitoring is crucial for addressing the global decline in insect populations. While most vision-based monitoring efforts analyze images of individual specimens, large-scale monitoring efforts yield "bulk images" where thousands of small insects are imaged in a single, high-resolution image. General object detection and instance segmentation models struggle to localize these insects due to the lack of discriminative visual features for tiny insects and the lack of relevant pretraining data. In this work, we present a case study that explores the effectiveness of super-resolution (SR) as a preprocessing step for tiny insect detection. We use the Mixed Arthropod Sample Segmentation and Identification (MassID45) dataset as a testbed for this task. MassID45 is the first dataset of its kind to feature high-resolution bulk images with instance segmentation and taxonomic classification labels for thousands of small, densely-packed insects. Our experiments show that the bilinear interpolation used in previous MassID45 baselines is suboptimal, and that applying more sophisticated upsampling methods boosts performance across multiple instance segmentation architectures. Leveraging several upsampling methods, ranging from bicubic interpolation to more sophisticated GAN and transformer-based SR models, we achieve an average precision score of 52.8\% on the MassID45 test set, representing an increase of 9.3 points from the previous baseline. These performance gains are most pronounced for small objects, proving that SR reconstructs visual details that aid in tiny object detection. Ultimately, our work establishes SR as an important step for bulk image analyses and automated biodiversity monitoring efforts.