TePD: Temporal Privileged Distillation for Amodal Counting Under Structured Occlusion

Despite the object's inclination to be unclear, which may take place through occlusion, vision models often need to infer images that do not have a sufficient amount of details of relevant objects in the scene. Presented is the concept of temporal privileged distillation, TePD, which represents a teacher-student approach used to acquire a hidden representation of the structure of the scene being analyzed. The technique involves using later images with unobstructed objects as a source of privileged knowledge. During the inference stage, TePD only needs the student data and the presently occluded image. We have assessed methodology using Cotton-TePD, the dataset featuring 3,000 pairs of UAV photos taken at two different instances before and after the chemical defoliation process. TePD’s latent cosine similarity value was found to equal 0.873, and the counting mean absolute error equaled 4.2, while MAE-ViT values were measured at 0.745 and 7.2, respectively. The results of the experiment were similar across three regions of the field, and the synthetic experiment devoid of textures yielded the results of 0.972 for its LCS. The ability of naturally occurring temporal revelation processes to serve as a means of training supervision for the occlusion-invariant visual representation is shown in the present findings without the need for synthetic occlusion masks or manually labeled amodal targets. This method is best used when occlusion decreases in proportion to time while the underlying target geometry maintains its stable configuration.