Quantifying the Impact of Occlusion Ratio and Type on Vision Model Detection Performance: Evaluating Robustness in Controlled Synthetic Occlusion Scenarios

Abstract

The increasing deployment of deep learning-based computer vision models in diverse applications demands functionality beyond ideal conditions. These models often face various forms of noise, one of the most prominent being different occlusions. This thesis examines occlusions by employing three isolated occluders (solid, perforated, volumetric) in a controlled synthetic environment, assessing their impact on a YOLO-based model trained in a similar environment by varying both occlusion ratio and camera viewpoint angles. The occlusion-related metrics were derived using two distinct methods, with occlusion type treated as a separate variable. Existing background work is also harnessed to support the research, which compares and contrasts existing findings with the experimental results, highlighting novelties that arise.

Analysis of the extracted metrics reveals a strong correlation between occlusion ratio and vision model performance, which can be quantified and modeled continuously rather than at discrete levels. Occlusion ratio is found to cause the performance of the model to follow a logistic-like collapse, with different occluder types and viewpoints producing distinct degradation profiles. While the findings both build on the reviewed studies and introduce novel insights, the narrow scope of the experiment limits the applicability of the results largely to an isolated, controlled environment. The experiment shows that occlusion is not a single-variable phenomenon, and should not be treated as such, though its practical applications outside the experimental environment remain limited. A broader study of a less limited nature is required in order to apply the insights beyond this environment, allowing for the evaluation of robustness of vision models in general under more realistic conditions.