Empirical Study of the Impact of Image Quality, Object Size, and Occlusion to Object Detection

EasyChair Preprint 9786

Abstract

Object detection is a crucial task in computer vision to identify and locate objects in an image or video. The performance of object detection algorithms is influenced by numerous factors such as image quality, object size, and occlusion. Image quality affects the clarity and resolution of the objects in the image, affecting the accuracy of object detection. Object size refers to the physical size of objects in an image and can influence their detection. Occlusion occurs when one object hides another, making it difficult to detect both objects accurately. This can result in false detections, missed detections, or a decrease in overall accuracy. Small objects also pose a significant challenge to object detection algorithms, as they can be difficult to detect and distinguish from their surroundings due to their limited size and distinctive features. Occlusion refers to the partial or complete hiding of an object by another object. It has a significant impact on the performance of object detection algorithms. This is because occlusion can make it difficult for the algorithms to detect and recognize objects, as well as to decide their position and orientation within an image. In this paper, our focuses are to find the effect of image quality, object size in an image, and object occlusion in an image. Our experimentation has revealed that the accuracy of our computer vision model is highly dependent on several factors, including image quality, occlusion, and object size. Specifically, we found that when presented with bad-quality images, the accuracy of our model drops significantly, sometimes plummeting as low as 25%. Similarly, we saw that heavily occluded images tend to decrease model accuracy, with our model achieving only 55% accuracy in such scenarios. In addition, we also noticed that our model struggled to achieve high accuracy with objects of varying sizes, resulting in low accuracy scores, with our model achieving only 67% accuracy in such scenarios.

Keyphrases: Box Blur, Gaussian Blur, Gaussian noise, Occlusion, computer vision, data-centric AI, object detection

@booklet{EasyChair:9786,
  author    = {Abhijit Kumar and Udit Mital and Ankit Gajera and Shivani Varanasi and Dibyendu Patra},
  title     = {Empirical Study of the Impact of Image Quality, Object Size, and Occlusion to Object Detection},
  howpublished = {EasyChair Preprint 9786},
  year      = {EasyChair, 2023}}