Search

Generative Adversarial Networks (GANs) are widely-used generative models for synthesizing complex and realistic data. However, mode collapse, where the diversity of generated samples is significantly lower than that of real samples, poses a major challenge for further applications. Our theoretical analysis demonstrates that the generator loss function is non-convex with respect to its parameters when there are multiple modes in real data. In particular, parameters that result in generated distributions with perfect partial mode coverage of the real distribution are the local minima of the generator loss function. To address mode collapse, we propose a unified framework called Dynamic GAN. This method detects collapsed samples in the generator by thresholding on observable discriminator outputs, divides the training set based on these collapsed samples, and trains a dynamic conditional model on the partitions. The theoretical outcome ensures progressive mode coverage and experiments on synthetic and real-world data sets demonstrate that our method surpasses several GAN variants. In conclusion, we examine the root cause of mode collapse and offer a novel approach to quantitatively detect and resolve it in GANs.

In this paper, we make an improvement on the conventional visual hull reconstruction method that runs on a single consumer graphics card. The target application is a high-precision diamond inclusion reconstruction system. One major contribution of this paper is an evaluation system of voxels for high-precision reconstruction. In contrast to existing approaches, it allows us to reconstruct the thin structure of the diamond inclusion. Based on this, we obtain a significant improvement in reconstruction precision, especially for thin inclusion.

The performance of feature learning for deep convolutional neural networks (DCNNs) is increasing promptly with significant improvement in numerous applications. Recent studies on loss functions clearly describing that better normalization is helpful for improving the performance of face recognition (FR). Several methods based on different loss functions have been proposed for FR to obtain discriminative features. In this paper, we propose an additive parameter depending on multiplicative angular margin to improve the discriminative power of feature embedding that can be easily implemented. In additive parameter approach, an automatic adjustment of the seedling element as the result of angular marginal seed is offered in a particular way for the angular softmax to learn angularly discriminative features. We train the model on publically available dataset CASIA-WebFace, and our experiments on famous benchmarks YouTube Faces (YTF) and labeled face in the wild (LFW) achieve better performance than the various state-of-the-art approaches.