Faster R-CNN with Generative Adversarial Occlusion Network for Object Detection

Abstract

Performing object detection on partially occluded objects is a challenging task due to the amount of variation in location, scale, and ratio present in real-world occlusion. A typical solution to this problem is to provide a large enough dataset with ample occluded samples for feature learning. However, this is rather costly given the amount of time and effort involved in the data collection process. In addition, even with such a dataset, there is no guarantee that it covers all possible cases of common occlusion in the real world. In this paper, we propose an alternate approach that utilizes the power of adversarial learning to reinforce the training of common object detection models. More specifically, we propose a Generative Adversarial Occlusion Network (GAON) capable of generating partially shaded training samples that are challenging for the object detector to classify. We demonstrate the efficacy of such an approach by conducting experiments on the Faster R-CNN detector, and the results indicate the superiority of our approach in improving the model's performance on occluded inputs.

DiffMIC: Dual-Guidance Diffusion Network for Medical Image Classification

“…Maximum-Mean Discrepancy (MMD) is to quantify the similarity between two distributions by comparing all of their moments [11,17]. It can be efficiently implemented using a kernel trick.…”

Class-imbalanced Unsupervised and Semi-Supervised Domain Adaptation for Histopathology Images

“…It estimates the distance between probability distributions of domains, encouraging low entropy and consistency on domain predictions [13]. • moment matching [14], [15] focuses on the moments of extracted features. Moment matching estimates one or more moments of features, estimating the discrepancy of feature distributions across domains to lessen the inconsistency between domains.…”

A Semi-Bayesian Nonparametric Hypothesis Test Using Maximum Mean Discrepancy with Applications in Generative Adversarial Networks

“…Moreover, Li et al (2015) considered k(X, Y) = T t=1 k G σ(t) (X, Y) in ( 2) and then used it to train the generator. Independently of Li et al (2015), Dziugaite et al (2015) proposed the same method to train G w but based on using a Bayesian optimization sequential design to set a value for σ and the number of neurons in each layer of network G w . Li et al (2017) proposed a learning technique by optimizing arg min…”

“…Previous work in applying MMD as a training criterion in GANs has only looked at the MMD as a test statistic in a frequentist two-sample test (Li et al, 2015;Dziugaite et al, 2015;Li et al, 2017;Bińkowski et al, 2018;Briol et al, 2019). However, these procedures are still liable to the drawbacks of frequentist testing.…”

“…Many frequentist GANs based on MMD can be found in the literature (Dziugaite et al, 2015;Briol et al, 2019;Li et al, 2015). However, BNP methods, while powerful, have comparatively not received much attention.…”