InSpect: A Curated Natural History Collection Dataset for Insect Specimen Understanding

Visual insect specimen understanding aims to identify biological taxa and characterize morphology-relevant structures from specimen images. Existing insect benchmarks have advanced this problem through large-scale pretraining and multimodal learning, but remain centered on taxonomic recognition with strong auxiliary modalities. We introduce \textbf{InSpect}, a curated natural history collection dataset containing 48,184 digitized insect specimen images with aligned crops, hierarchical taxonomy, label-derived structured metadata, and 1,859 images with fine-grained anatomical part masks. We thus establish two benchmark settings. The first is \textbf{open taxonomic recognition}, which evaluates zero-shot, fine-tuned, and unseen-taxon recognition using cropped insect images, label-derived metadata, or their combination. The second is \textbf{fine-grained anatomical part segmentation}, which evaluates supervised and text-guided segmentation of insect structures such as antennae, legs, wings, and body regions. Experiments show that open taxonomic recognition remains challenging despite strong closed-set performance, and that specimen metadata provides useful but unstable contextual cues. For anatomical segmentation, closed-set segmentation models remain weak on thin structures, and zero-shot open-vocabulary models often fail to distinguish insect parts from the body or background. Together, these results show that InSpect enables systematic evaluation of insect specimen understanding across cropped visual morphology, label-derived context, and fine-grained anatomical annotations.