This work presents a method for volumetric seismic data compression by coupling a 3-D convolution-based autoencoder to a generative adversarial network (GAN). The main challenge of 3-D convolutional autoencoders for data compression is how to fully exploit volumetric redundancy while keeping reasonable latent representation dimensions. Our method is based on a convolutional neural network for seismic data compression called 3DSC. Its encoder and decoder use 3-D convolutions and are connected by a latent representation with the same dimensions as its 2-D network counterparts. Our main hypothesis is that the 3DSC architecture can be improved by adversarial training. We, thus, propose a new 3-D-based seismic data compression method (3DSC-GAN) by coupling the 3DSC network to a GAN. The seismic data decoder is used as a generator of poststack data that are integrated with a discriminator module to better exploit 3-D redundancy. Results show that our method outperforms previous seismic data compression methods for very low target bit rates, increasing the peak signal-to-noise ratio (PSNR) with fairly high visual quality.

We approach the problem of 3-D poststack seismic data compression by training a model based on a deep autoencoder. Our network architecture is trained to consider the similarity between 3-D seismic sections drawn from one or multiple seismic volumes. A whole seismic volume is compressed with the latent representations of each of its composing volumetric sections. The goal is to compress the seismic data at very low bit rates with high-quality reconstruction. Our model is suitable for training general compressors from multiple seismic surveys or for specialized compression of a single seismic volume. Results show that our method can compress seismic data with extremely low bit rates, below 0.3 bits-per-voxel (bpv) while yielding peak signal-to-noise ratio (PSNR) values over 40 dB.

Abstract We present a method to recover the 3D flying shape of a sail using passive markers. In the navigation and naval architecture domain, retrieving the sail shape may be of immense value to confirm or contest simulation results, and to aid the design of new optimal sails. Our acquisition setup is very simple and low-cost, as it is only necessary to fix a series of printable markers on the sail and register the flying shape in real sailing conditions from a side vessel with a single camera. We reconstruct the average sail shape during an interval where the sailor maintains the sail as stable as possible. The average is further improved by a Bundle Adjustment algorithm. We tested our method in a real sailing scenario and present promising results. Quantitatively, we show the precision in regards to the reconstructed markers area and the reprojected points. Qualitatively, we present feedback from domain experts who evaluated our results and confirmed the usefulness and quality of the reconstructed shape.

Recent deep learning techniques have achieved satisfactory results for various image-related problems. However, many research questions remain open in tasks involving video sequences. Several applications demand the understanding of complex events in videos, such as traffic monitoring, person re-identification, security and surveillance. In this work, we address the problem of human action recognition in videos through a multi-stream network that incorporates both spatial and temporal information. The main contribution of our work is a stream based on a new variant of the visual rhythm, called Learnable Visual Rhythm (LVR). We employ a deep network to extract features from the video frames in order to generate the rhythm. The features are collected at multiple depths of the network to enable the analysis of different abstraction levels. This strategy significantly outperforms the handcrafted version on the UCF101 and HMDB51 datasets. Experiments conducted on these datasets show that our final multi-stream network achieved competitive results compared to state-of-the-art approaches.

Advances in digital technology have increased event recognition capabilities through the development of devices with high resolution, small physical dimensions and high sampling rates. The recognition of complex events in videos has several relevant applications, particularly due to the large availability of digital cameras in environments such as airports, banks, roads, among others. The large amount of data produced is the ideal scenario for the development of automatic methods based on deep learning. Despite the significant progress achieved through image-based deep networks, video understanding still faces challenges in modeling spatio-temporal relations. In this work, we address the problem of human action recognition in videos. A multi-stream network is our architecture of choice to incorporate temporal information, since it may benefit from pre-trained deep networks for images and from handcrafted features for initialization. Furthermore, its training cost is usually lower than video-based networks. We explore visual rhythm images since they encode longer-term information when compared to still frames and optical flow. We propose a novel method based on point tracking for deciding the best visual rhythm direction for each video. Experiments conducted on the challenging UCF101 and HMDB51 data sets indicate that our proposed stream improves network performance, achieving accuracy rates comparable to the state-of-the-art approaches.

On the problem of tracking objects in videos, a recent and distinguished approach combining tracking and detection methods is the TLD framework. The detector identifies the object by its supposedly confirmed appearances. The tracker inserts new appearances into the model using apparent motion. Their outcomes are integrated by using the same similarity metric of the detector which, in our point of view, leads to biased results. We propose a mediator method to integrate the motion tracker and detector by combining their estimations. Our results show that when the mediaton strategy is independent of both tracker/detector metrics, the overall tracking is improved for objects with high appearance variations throughout the video.

We propose a method for the detection of high frequency regions using multiresolution analysis and orientation tensors. A scalar field representing multiresolution edges is obtained. Local maxima of this scalar space indicate regions having coincident detail vectors in multiple scales of a wavelet decomposition. This is useful for finding edges, textures, collinear structures and salient regions for computer vision methods. The image is decomposed into several scales using the discrete wavelet transform (DWT). The resulting detail spaces form vectors indicating intensity variations which are adequately combined using orientation tensors. The multivariate data of the resulting tensor field provides fair estimations of high frequency regions. Using these tensors, a positive scalar is computed for each original image pixel. Our results show that this descriptor indicates areas having relevant intensity variation in multiple scales.

We describe a new technique that uses active scene illumination to perform foreground-background segmentation and recover partial HDR information. We explore the fact that relative tones can be recovered by varying illumination intensity, without knowing the camera response function. In our approach, the scene is illuminated with an uncalibrated projector and two images of the scene are captured under different illumination conditions. By taking advantage of the fact that the projector can be set up to illuminate only the foreground, we are able to distinguish the foreground from the background. The output of our system is a segmentation mask, together with a image with additional tonal information for the foreground pixels. As an application, we show how to produce spatially variant tone mapped images, where background and foreground receive different treatments. The segmentation and the visualization algorithms are implemented in real-time, and can be used to produce range-enhanced video sequences.

Abstract We propose a new strategy to estimate surface normal information from highly noisy sparse data. Our approach is based on a tensor field morphologically adapted to infer normals. It acts as a three‐dimensional structuring element of smooth surfaces. Robust orientation inference for all input elements is performed by morphological operations using the tensor field. A general normal estimator is defined by combining the inferred normals, their confidences and the tensor field. This estimator can be used to directly reconstruct the surface or give input normals to other reconstruction methods. We present qualitative and quantitative results to show the behavior of the original methods and ours. A comparative discussion of these results shows the efficiency of our propositions.