We introduce GaussianOcc, a systematic method that investigates the two usages of Gaussian splatting for fully self-supervised and efficient 3D occupancy estimation in surround views. First, traditional methods for self-supervised 3D occupancy estimation still require ground truth 6D poses from sensors during training. To address this limitation, we propose Gaussian Splatting for Projection (GSP) module to provide accurate scale information for fully self-supervised training from adjacent view projection. Additionally, existing methods rely on volume rendering for final 3D voxel representation learning using 2D signals (depth maps, semantic maps), which is both time-consuming and less effective. We propose Gaussian Splatting from Voxel space (GSV) to leverage the fast rendering properties of Gaussian splatting. As a result, the proposed GaussianOcc method enables fully self-supervised (no ground truth pose) 3D occupancy estimation in competitive performance with low computational cost (2.7 times faster in training and 5 times faster in rendering). The relevant code is available in https://github.com/GANWANSHUI/GaussianOcc.git.

Monocular depth estimation (MDE) is crucial for various computer vision applications, but existing methods often struggle to balance inference speed and accuracy when processing large-region visual information. This paper introduces LR2Depth, a novel MDE method that addresses this challenge by utilizing large-kernel convolution on low-resolution feature maps for efficient large-region feature aggregation. Our approach leverages the fact that each pixel on low-resolution feature maps corresponds to a larger region of the original image, allowing for fast and accurate depth predictions at a lower inference cost. Extensive experiments on NYU-Depth-V2, KITTI, and SUN RGB-D datasets demonstrate that LR2Depth not only achieves state-of-the-art performance but also operates approximately twice as fast as previous MDE methods. Notably, at the time of submission, LR2Depth secured the top-1 position on the KITTI depth prediction online benchmark. The code is available in the project page.

Text-driven style transfer for Neural Radiance Fields (NeRFs) is an emerging research topic that leverages text descriptions instead of reference style images to apply style transfer. However, existing methods for stylizing NeRFs predominantly struggle to extend to 4D dynamic scenes, due to NeRFs' inherent limitation to static environments. Moreover, these current methods require training for each specific text input, which limits them to a single style description and significantly hampers generalizability and applications. In this paper, we introduce a novel approach to zero-shot text-driven 4D style transfer that adopts text inputs into the CLIP's style space with a canonical feature volume. Specifically, using geometric priors from pre-trained dynamic Neural Radiance Fields, we train a canonical feature volume by rendering feature maps under the supervision of a pre-trained VGG encoder. Then we utilize CLIP's multi-modal embedding to connect the text descriptions with style images and learn a canonical style transformation matrix in CLIP's feature space. Experiments show that our method achieves zero-shot text-driven style transfer for dynamic neural radiance fields and maintains good multi-view and cross-time consistency.

Neural radiance fields have made a remarkable breakthrough in the novel view synthesis task at the 3D static scene. However, for the 4D circumstance (e.g., dynamic scene), the performance of the existing method is still limited by the capacity of the neural network, typically in a multilayer perceptron network (MLP). In this article, we utilize 3D Voxel to model the 4D neural radiance field, short as V4D, where the 3D voxel has two formats. The first one is to regularly model the 3D space and then use the sampled local 3D feature with the time index to model the density field and the texture field by a tiny MLP. The second one is in look-up tables (LUTs) format that is for the pixel-level refinement, where the pseudo-surface produced by the volume rendering is utilized as the guidance information to learn a 2D pixel-level refinement mapping. The proposed LUTs-based refinement module achieves the performance gain with little computational cost and could serve as the plug-and-play module in the novel view synthesis task. Moreover, we propose a more effective conditional positional encoding toward the 4D data that achieves performance gain with negligible computational burdens. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance at a low computational cost.

In the stereo matching task, the 3D convolution network can effectively aggregate the cost volume with the strong representation ability to model the spatial and depth dimensions but with the disadvantage of a high computational cost. In this letter, we revisit the 3D convolution network and its common variant, and then propose the Depth Shift Module (DSM) to model the cost volume in the depth dimension which could imitate the 3D convolution function with the computational complexity of the 2D convolution. The proposed DSM is easy to extend to present 3D cost aggregation methods in stereo matching with less inference time, lower computational complexity, and minor precision loss. Moreover, a novel compact but efficient stereo matching framework named HybridNet is proposed. This framework can hybridize the 2D convolution layer with the proposed DSM to effectively aggregate the cost volume. The proposed HybridNet achieves a better trade-off between the performance, computational complexity, and model size ( e.g. , 30% less than the size of AANet and 25% less than the size of PSMNet) in public open-source datasets ( e.g. , Scene Flow and KITTI Stereo 2015). The relevant code is available at https://github.com/GANWANSHUI/HybridNet .

In this paper, a computation efficient regression framework is presented for estimating the 6D pose of rigid objects from a single RGB-D image, which is applicable to handling symmetric objects. This framework is designed in a simple architecture that efficiently extracts point-wise features from RGB-D data using a fully convolutional network, called XYZNet, and directly regresses the 6D pose without any post refinement. In the case of symmetric object, one object has multiple ground-truth poses, and this one-to-many relationship may lead to estimation ambiguity. In order to solve this ambiguity problem, we design a symmetry-invariant pose distance metric, called average (maximum) grouped primitives distance or A(M)GPD. The proposed A(M)GPD loss can make the regression network converge to the correct state, i.e., all minima in the A(M)GPD loss surface are mapped to the correct poses. Extensive experiments on YCB-Video and TLESS datasets demonstrate the proposed framework's substantially superior performance in top accuracy and low computational cost. The relevant code is available in https://github.com/GANWANSHUI/ES6D.git.