Recent years have witnessed the remarkable success of Text-to-3D generation, particularly with the rise of mainstream conditional diffusion models (DMs). Though achieving substantial progress, existing methods still face a knotty "human preference" dilemma, that is the 3D contents generated by the models often deviate greatly from the desired effects (e.g., perspective, aesthetics, shading, appearance, etc.) due to the lack of attention to human preferences. To mitigate the limitation of data deficiency and enable human preference learning, we first elaborately curate the HP3D, a text-to-3D dataset with expert preference annotations which is initally captioned by the multimodal large model LLava and then refined by human expert. Based on such a brand-new HP3D, we further propose DreamAlign, a reward-free method that does not require designing any complex reward models whereas only by introducing a light-weight lora adapter and then designing a novel direct 3D preference optimization (D-3DPO) algorithm for training. Moreover, in the stage of text-to-3D we design an additional Preference Contrastive Feedback training for score distillation sampling, which enables the generated 3D objects to align the human preferences (e.g., aesthetics, material, etc.). Extensive experiments demonstrate that DreamAlign consistently achieves state-of-the-art performance on generative effects and human preference alignment across various benchmark evaluations.

In recent years, significant breakthroughs have been made in audio-guided 3D facial animation. However, existing methods mainly focus on lip shape and audio consistency and still face key challenges to achieve alignment between facial emotions and speech emotions. To overcome this limitation, we introduce EmoRLTalk, a novel framework that integrates offline reinforcement learning to implicitly capture the intricate relationship between 3D facial landmarks and blendshape parameters, thereby enhancing the granularity of emotional expression. Furthermore, we harness the strength of conditional diffusion models to synthesize facial motions that are emotionally coherent with the input speech. Additionally, based on the multi-task learning paradigm, we construct a collaborative training framework of a regression main task and a classification sub-task. Specifically, we use emotion classification of blendshape as a sub-task to further improve the model’s ability to express facial emotions. To further enhance system controllability, we integrate the ControlNet module, allowing users to achieve precise facial expression control. Extensive experiments demonstrate that EmoRLTalk achieves superior emotional expressiveness and lip-sync performance compared to previous approaches.

Traditional emotion recognition in conversation (ERC) studies are usually designed to predict a fixed set of predetermined emotion categories. This limited supervision diminishes the expressive power of the data, resulting in failure to capture the complexity of human emotions in conversation. Learning from a well-designed fine-grained representation of emotions offers a promising alternative that utilizes a wider range of supervision. In this paper, the proposed Fine-grained Multidimensional Emotion Representation Learning (FMERL) framework integrates multitask learning and contrastive learning, and extends the emotion representation of valence, arousal and dominance (VAD) from the psychological field to both continuous and discrete forms. Firstly, the emotion features from text, audio, and visual modalities are extracted. Then, the multimodal features are fused by a transformer-based model. The multitask learning module consists of three networks: the valence network, arousal network, and dominance network, for learning the continuous fine-grained emotion representations from the fused multimodal features. The contrastive learning aligns fused multimodal features with the discrete fine-grained emotion representations derived through prompt engineering applied to a large language model. The transferable ability of contrastive learning enables FMERL to map the semantic information of emotion representation and fused multimodal features into a shared embedding space, thereby understanding their semantic relationships and enabling zero-shot learning for unseen emotion classes. Experimental results on the IEMOCAP and MELD datasets have shown that FMERL achieves state-of-the-art performance in emotion recognition and implements zero-shot learning.

As one of the most popular and sought-after generative models in recent years, diffusion models have sparked the interests of many researchers and steadily shown excellent advantage in various generative tasks such as image synthesis, video generation, bioinformatics engineering, 3D scene rendering and multimodal generation, relying on their dense theoretical principles and reliable application practices. The remarkable success of these recent efforts on diffusion models comes largely from progressive design principles and efficient architecture, training, inference, and deployment methodologies. However, there has not been a comprehensive and in-depth review to summarize these principles and practices to help the rapid understanding and application of diffusion models. In this survey, we provide a new efficiency-oriented perspective on these existing efforts, which mainly focuses on the profound principles and efficient practices in architecture designs, model training, fast inference and reliable deployment, to guide further theoretical research, algorithm migration and model application for new scenarios in a reader-friendly way.

First, since time consumption for extracting$\alpha$will become quite tedious for very large images by pixelwise eigendecomposition, we proposed a fast$\alpha$angle's solution. Second,$\alpha$may be not unique that results in the invalidity of$\mathrm{H}/\ \alpha$decomposition. So we early proposed a sound discriminant to distinguish$\alpha$is unique or not. Third, since depolarization is serious and$\alpha$is unstable in high entropy zone, grass and flourishing canopy etc may be misclassified as double bounce scattering. Therefore we proposed a fast algorithm that$\mathrm{H}/\alpha$decomposition is combined with the model-based decomposition, which overcomes the shortcomings of the invalidity of$\mathrm{H}/\alpha$decomposition and the misclassification in high-entropy zone, and its time consumption is much shorter than$\mathrm{H}/\alpha$decomposition.

Fuzzy regression model is developed to construct the relationship between independent variable and dependent variable in a fuzzy environment. In order to increase the explanatory performance of fuzzy regression model, the least-squares method usually is applied to determine the numeric coefficients based on the concept of distance. In this paper, we consider the fuzzy linear regression model with fuzzy input, fuzzy output and crisp parameters and introduce a new distance based on the geometric centroid and incentre points (GCIP) of triangular fuzzy number, merge least-squares method with the new GCIP distance and propose least-squares GCIP distance method. Finally, an example of employee job performance is given to illustrate the effectiveness and feasibility of the method. Comparisons with existing methods show that total estimation error using the same distance criterion, the explanatory performance of the GCIP method is satisfactory, and the calculation is relatively simple.

Kernel methods with specifically designed kernel function are suitable for dealing with practical nonlinear problems. However, kernel methods have found limited applications to synthetic aperture radar (SAR) image change detection in that their performances are affected by the inherent multiplicative speckle noise of SAR images. It is known that the spatial-contextual information is helpful in suppressing the degrading effects of the noise. Therefore, a label-information composite kernel (LIC kernel) constructed on the basis of the spatial-contextual information is proposed in this paper for SAR image change detection. A typical spatial information, the output-space label-neighborhood information that is extracted using all labels in the neighborhood of each pixel, may enhance noise immunity, but with inaccurate edge locations simultaneously. Consequently, the anisotropic Gaussian kernel model is utilized for analyzing anisotropic textures of the bitemporal images, and then, a comparison scheme acting on the input-space textures of the bi-temporal images is proposed to supervise the extraction of the output-space label-neighborhood information in the construction of the LIC kernel. The constructed LIC kernel is of good preservation of edge locations of changed areas as well as strong noise immunity. The LIC kernel is updated iteratively with the newest change map outputted from the support vector machine, until the change map converges. Experiments on real SAR images demonstrate the effectiveness of the LIC kernel method and illustrate that it has both strong noise immunity and good preservation of edge locations of changed areas for SAR image change detection.

The triplet Markov field (TMF) can obtain more promising classification results of nonstationary images than the Markov random field (MRF). However, TMF has limitedly specialized applications to polarimetric synthetic aperture radar (PolSAR) images with nonstationarity properties. In addition, it is difficult to interpret the meaning of the auxiliary field derived by TMF. This implies that the auxiliary field may not have the physical meaning. We propose Wishart TMF with a specific auxiliary field for PolSAR image classification. We define a smoothness characteristic, which describes the extent of pixel smoothness in its neighborhood. This characteristic acts on the energy of the proposed TMF to supervise the classification of the auxiliary field. The auxiliary field can distinguish the smoothness stationarity and nonsmoothness stationarity of PolSAR images, which indicates that the auxiliary field has the specific physical meaning. The effectiveness of the proposed TMF is demonstrated by real PolSAR image classification experiments.

The motivation of this letter is to resolve the nonnegative eigenvalue constraint (NNEC) problem of four-component decomposition (FCD). It is analyzed that the NNEC is an essential requirement for remainder coherency matrices in the FCD, however the measured polarimetric synthetic aperture radar (POLSAR) data experiment shows there exits the NNEC problem that some remainder coherency matrices of the FCD do not satisfy the NNEC, which means these matrices are not positive semi-definite. In addition, it is analyzed that the scheme using the nonnegative eigenvalue decomposition (NNED) for three-component decomposition (TCD) cannot be directly extended to the FCD to overcome the NNEC problem, so a scheme using the NNED for the FCD is proposed as follow. From matrix theory, we draw a conclusion that if the last remainder coherency matrix satisfies the NNEC, then all remainder coherency matrices also satisfy the NNEC; we successively analyze that the NNEC problem of the last remainder coherency matrices results from the overestimation of scattering powers. Then a shrinkage coefficient is used to depress all possible overestimations of scattering powers, and the overestimation case with the minimum remainder power is chosen to resolve the NNEC problem. Moreover, we have simplified the solution to NNED, which is used to calculate the shrinkage coefficient. The measured POLSAR data experiment shows that the proposed FCD can further enhance double-bounce scattering and depress volume scattering for urban areas.

The triplet Markov field (TMF) model recently proposed is suitable for tackling the nonstationary image segmentation. In this letter, we propose a hierarchical TMF (HTMF) model for unsupervised synthetic aperture radar (SAR) image segmentation. In virtue of the Bayesian inference on the quadtree, the HTMF model captures the global and local image characteristics more precisely in the bottom-up and top-down probability computations. In this way, the underlying spatial structure information is effectively propagated. To model the SAR data related to radar backscattering sources, generalized Gamma distribution is utilized. The effectiveness of the proposed HTMF model is demonstrated by application to simulated data and real SAR image segmentation.

The Microwave Radiation Imager (MWRI) on board the FengYun-3A/B satellites observes the Earth atmosphere at 10.65, 18.7, 23.8, 36.5, and 89.0 GHz with each having dual polarization. Its calibration system is uniquely designed with a main reflector viewing both cold and hot calibration targets. Two quasi-optical reflectors are used to reflect the radiation from the hot load and cold space to the main reflector. In the MWRI calibration process, a radiation loss in the beam transmission path must be taken into account. The loss factor in the hot load transmission path is derived using the antenna pattern data measured on ground and satellite data observing over the Amazon forest where the scene temperature is steady and close to the hot load. The instrument nonlinearity factors at different channels are also evaluated over a wide range of brightness temperatures and compared with the results from the ground vacuum test. After a cross-calibration with Windsat data, atmospheric products are derived from MWRI brightness temperatures with the accuracy similar to those from the legacy sensors (e.g., the Special Sensor Microwave/Imager).