Point clouds are essential for storage and transmission of 3D content. As they can entail significant volumes of data, point cloud compression is crucial for practical usage. Recently, point cloud geometry compression approaches based on deep neural networks have been explored. In this paper, we evaluate the ability to predict perceptual quality of typical voxel-based loss functions employed to train these networks. We find that the commonly used focal loss and weighted binary cross entropy are poorly correlated with human perception. We thus propose a perceptual loss function for 3D point clouds which outperforms existing loss functions on the ICIP2020 subjective dataset. In addition, we propose a novel truncated distance field voxel grid representation and find that it leads to sparser latent spaces and loss functions that are more correlated with perceived visual quality compared to a binary representation. The source code is available at <uri>https://github.com/mauriceqch/2021_pc_perceptual_loss</uri>.
The last decades witnessed an increasing number of works aiming at proposing objective measures for media quality assessment, i.e. determining an estimation of the mean opinion score (MOS) of human observers. In this contribution, we investigate a possibility of modeling and predicting single observer’s opinion scores rather than the MOS. More precisely, we attempt to approximate the choice of one single observer by designing a neural network (NN) that is expected to mimic that observer behavior in terms of visual quality perception. Once such NNs (one for each observer) are trained they can be looked at as “virtual observers” as they take as an input information about a sequence and they output the score that the related observer would have given after watching that sequence. This new approach allows to automatically get different opinions regarding the perceived visual quality of a sequence whose quality is under investigation and thus estimate not only the MOS but also a number of other statistical indexes such as, for instance, the standard deviation of the opinions. Large numerical experiments are performed to provide further insight into a suitability of the approach.
Recent advances in deep learning (DL) have led to great success in tasks of computer vision and pattern recognition. Sharing pre-trained DL models has been an important means to promote the rapid progress of research community and development of DL based systems. However, it also raises challenges to model authentication. It is quite necessary to protect the ownership of the DL models to be released. In this paper, we present a digital watermarking technique to deep neural networks (DNNs). We propose to mark a DNN by inserting an independent neural network that allows us to use selective weights for watermarking. The independent neural network is only used in the training phase and watermark verification phase, and will not be released publicly. Experiments have shown that, the performance of marked DNN on its original task will not be degraded significantly. Meantime, the watermark can be successfully embedded and extracted with a low neural network loss even under the common attacks including model fine-tuning and compression, which has shown the superiority and applicability of the proposed work.