Dinei Florencio's research while affiliated with Microsoft and other places

... In recent years, machine learning techniques have also gained popularity in the selection of near-field beamforming sensors. These methods take advantage of the power of data-driven algorithms to learn patterns and relationships from training data, enabling the identification of sensors that provide the most beamforming information [25]. ...

... Due to high similarities between object and background, COD is usually more challenging than traditional object segmentation tasks. Early approaches to COD used different handcrafted features like texture and contrast 21 , 3d-convexity 22 , optical flow 23 or try to solve the problem in the frequency domain 24 . As the complexity of the scene rises, these approaches often struggle to segment the camouflaged objects. ...

... This technique possesses the coding of an interlayer residual feature. In their study, De et al. [23] established context by measuring the distances between the actual voxels and those of a reference PC. They then used this context to determine the probabilities of the node states, which helped lower the bit rate. ...

... Neural networks can learn any function given a sufficient number of parameters and training data, and have been successfully used in fields like image classification, inverse problems, image segmentation, [14][15][16][17] and many fields in acoustics. 18 They have also already been used in beamforming, mostly in areas related to speech recognition for cleaning speech recordings, 19 and in recent years, intensively in various approaches for sound source localization. 20 Related to beamforming, some researchers have successfully applied neural networks to estimate Capon beamforming weights, 21 where the neural network indirectly learned the optimal weightings. ...

... Visual "capture" effects, where visual stimuli associated with a sound source affect its localization in space, have been extensively studied in experimental psychology [35]. Similar effects were recently observed in VR using generic HRTFs [9]. Thus, in future works, we plan to investigate the effect of different approaches and degrees of HRTF personalization, as well as to compare our acoustic simulation approach to other auralisation engines and simulators, such as Project Triton [58], in relation to the level of HRTF personalization. ...

... Other tools that implement supervised conditional GAN include [104] [169] [170] [93]. In [171], Bayesian network is exploited to generate estimated clean speech from a noisy one. ...

... For the computation of the vector of beamformer coefficients f k , several criteria exist. Here f k is estimated using a MSE based criterion similar to [22] f k = argmin ...

... In initial investigations, the majority of approaches employ basic features such as texture, edges, luminosity, and color to differentiate the camouflaged object from its surroundings [36][37][38][39][40][41]. Nevertheless, camouflage often disrupts the inherent features to deceive the observer, rendering these approaches comparatively less efficacious. ...

... Given the above discussion and the current experimental limitations, we plan to apply our methodology along four main directions to evaluate the influence on co-immersion of: (a) ERs simplification following previous related works [15,31]. To this end, a possible artificial reverberator are Scattering Delay Networks (SDN) [18] which model ERs according to physical properties; (b) different listening environments, considering a wider set than the three VAEs used in the present study; (c) different conversational scenarios, considering concurrent talking [30], turn-taking dialogues, or partially overlapping speakers; (d) the effect of visual elements in modulating auditory perception and cognition, e.g. visual rendering matching the acoustics features (size, material, etc.). ...

... It showed a better performance in comparison with interpolation methods without a priori knowledge. In recent years, more novel interpolation methods using prior data to model HRTFs have been proposed, such as a Gaussian process regression method (Luo et al., 2013), a manifold learning method (Grijalva et al., 2017), a spherical-harmonics-based spatial aliasing cancellation method (Alon et al., 2018), etc. However, the above methods are relatively complicated, and no subjective experiments have been conducted to evaluate the perceptual errors. ...