Semantic Decoupling based Semantic Scene Completion from a Single Depth Image

As a voxel-wise labeling task, semantic scene completion (SSC) tries to simultaneously infer the occupancy and semantic labels for a scene from a single depth and/or RGB image. The key challenge for SSC is how to effectively take advantage of the 3D context to model various objects or stuffs with severe variations in shapes, layouts, and visibility. To handle such variations, we propose a novel module called anisotropic convolution, which properties with flexibility and power impossible for the competing methods such as standard 3D convolution and some of its variations. In contrast to the standard 3D convolution that is limited to a fixed 3D receptive field, our module is capable of modeling the dimensional anisotropy voxel-wisely. The basic idea is to enable anisotropic 3D receptive field by decomposing a 3D convolution into three consecutive 1D convolutions, and the kernel size for each such 1D convolution is adaptively determined on the fly. By stacking multiple such anisotropic convolution modules, the voxel-wise modeling capability can be further enhanced while maintaining a controllable amount of model parameters. Extensive experiments on two SSC benchmarks, NYU-Depth-v2 and NYUCAD, show the superior performance of the proposed method.

Semantic Scene Completion (SSC) aims to reconstruct complete 3D scenes with precise voxel-wise semantics from the single-view incomplete input data, a crucial but highly challenging problem for scene understanding. Although SSC has seen significant progress due to the introduction of 2D semantic priors in recent years, the occluded parts, especially the rear-view of the scenes, are still poorly completed and segmented. To ameliorate this issue, we propose a novel deep learning framework for 3D SSC, named Planar Convolution and Attention-based Network (PCANet), to effectively extend high-precision predictions of the front-view surface to the rear-view occluded areas. Specifically, we decompose the traditional convolutional layer into three successive planar convolutions to form a Planar Convolution Residual (PCR) block, which maintains the planar features of the 3D scene. Afterward, the Planar Attention Module (PAM) is proposed to capture three different planar attentions and harvest the global context from the front surface to the rear occluded areas to improve the overall accuracy. Extensive experiments on the real NYU and NYUCAD datasets and the synthetic SUNCG-RGBD dataset demonstrate that our proposed framework can generate high-quality SSC results in both front and rear views and outperforms the state-of-the-art approaches trained in an end-to-end manner without additional data.

We study outdoor 3D scene understanding, a challenging task demanding the intelligent system to infer both geometry and semantics from a single-view image – a critical skill for autonomous vehicles to navigate in the real 3D world. Towards this end, we present an instance-aware monocular semantic scene completion framework. To the best of our knowledge, this is the first endeavor specifically targeting the challenge of instance perception in the camera-based semantic scene completion task. Our method consists of two stages. In stage I, we design a region-based VQ-VAE network, providing an effective solution for 3D occupancy prediction. In stage II, we first introduce an instance-aware attention module, explicitly incorporating instance-level cues captured from mask images to enhance the instance features in RGB images. Then we leverage the deformable cross-attention to aggregate image features corresponding to each voxel query and utilize the deformable self-attention to refine query proposals. We combine these key ingredients and evaluate our method on two challenging datasets, namely SemanticKITTI and SSCBench-KITTI-360. The results unequivocally demonstrate the superiority of our proposed method over the state-of-the-art VoxFormer-S. Specifically, our method surpasses VoxFormer-S by 0.22 IoU and 0.72 mIoU on the validation set and achieves an impressive improvement of 3.04 IoU and 1.06 mIoU on the SSCBench-KITTI-360 validation set. Meanwhile, our approach ensures accurate perception of critical instances, thereby exhibiting its exceptional performance and potential for practical deployment.

We tackle the practical problem of hand pose estimation from a single noisy depth image. A dedicated three-step pipeline is proposed: Initial estimation step provides an initial estimation of the hand in-plane orientation and 3D location, Candidate generation step produces a set of 3D pose candidate from the Hough voting space with the help of the rotational invariant depth features, Verification step delivers the final 3D hand pose as the solution to an optimization problem. We analyze the depth noises, and suggest tips to minimize their negative impacts on the overall performance. Our approach is able to work with Kinect-type noisy depth images, and reliably produces pose estimations of general motions efficiently (12 frames per second). Extensive experiments are conducted to qualitatively and quantitatively evaluate the performance with respect to the state-of-the-art methods that have access to additional RGB images. Our approach is shown to deliver on par or even better results.

Panoptic Segmentation aims to provide an understanding of background (stuff) and instances of objects (things) at a pixel level. It combines the separate tasks of semantic segmentation (pixel level classification) and instance segmentation to build a single unified scene understanding task. Typically, panoptic segmentation is derived by combining semantic and instance segmentation tasks that are learned separately or jointly (multi-task networks). In general, instance segmentation networks are built by adding a foreground mask estimation layer on top of object detectors or using instance clustering methods that assign a pixel to an instance center. In this work, we present a fully convolution neural network that learns instance segmentation from semantic segmentation and instance contours (boundaries of things). Instance contours along with semantic segmentation yield a boundary aware semantic segmentation of things. Connected component labeling on these results produces instance segmentation. We merge semantic and instance segmentation results to output panoptic segmentation. We evaluate our proposed method on the CityScapes dataset to demonstrate qualitative and quantitative performances along with several ablation studies. Our overview video can be accessed from https://youtu.be/wBtcxRhG3e0.

Chapter 12 - Semantic scene segmentation for robotics

Geometry-semantic aware for monocular 3D Semantic Scene Completion

Content-based 3D shape retrieval is an important problem in computer vision. Traditional retrieval interfaces require a 2D sketch or a manually designed 3D model as the query, which is difficult to specify and thus not practical in real applications. With the recent advance in low-cost 3D sensors such as Microsoft Kinect and Intel Realsense, capturing depth images that carry 3D information is fairly simple, making shape retrieval more practical and user-friendly. In this paper, we study the problem of cross-domain 3D shape retrieval using a single depth image from low-cost sensors as the query to search for similar human designed CAD models. We propose a novel method using an ensemble of autoencoders in which each autoencoder is trained to learn a compressed representation of depth views synthesize d from each database object. By viewing each autoencoder as a probabilistic model, a likelihood score can be derived as a similarity measure. A domain adaptation layer is built on top of autoencoder outputs to explicitly address the cross-domain issue (between noisy sensory data and clean 3D models) by incorporating training data of sensor depth images and their category labels in a weakly supennsed learning formulation. Experiments using real-world depth images and a large-scale CAD dataset demonstrate the effectiveness of our approach, which offers significant improvements over state-of-the-art 3D shape retrieval methods.

We propose ESSC-RM, a plug-and-play Enhancing framework for Semantic Scene Completion with a Refinement Module, which can be seamlessly integrated into existing semantic scene completion (SSC) models. ESSC-RM operates in two phases: a baseline SSC network first produces a coarse voxel prediction, which is subsequently refined by a 3D U-Net-based Prediction Noise-Aware Module (PNAM) and Voxel-level Local Geometry Module (VLGM) under multiscale supervision. Experiments on SemanticKITTI show that ESSC-RM consistently improves semantic prediction performance. When integrated into CGFormer and MonoScene, the mean IoU increases from 16.87 to 17.27% and from 11.08 to 11.51%, respectively. These results demonstrate that ESSC-RM serves as a general refinement framework applicable to a wide range of SSC models. Project page: https://github.com/LuckyMax0722/ESSC-RM and https://github.com/LuckyMax0722/VLGSSC.

The field of continual deep learning is an emerging field and a lot of progress has been made. However, concurrently most of the approaches are only tested on the task of image classification, which is not relevant in the field of intelligent vehicles. Only recently approaches for class-incremental semantic segmentation were proposed. However, all of those approaches are based on some form of knowledge distillation. At the moment there are no investigations on replay-based approaches that are commonly used for object recognition in a continual setting. At the same time while unsupervised domain adaption for semantic segmentation gained a lot of traction, investigations regarding domain-incremental learning in an continual setting is not well-studied. Therefore, the goal of our work is to evaluate and adapt established solutions for continual object recognition to the task of semantic segmentation and to provide baseline methods and evaluation protocols for the task of continual semantic segmentation. We firstly introduce evaluation protocols for the class- and domain-incremental segmentation and analyze selected approaches. We show that the nature of the task of semantic segmentation changes which methods are most effective in mitigating forgetting compared to image classification. Especially, in class-incremental learning knowledge distillation proves to be a vital tool, whereas in domain-incremental learning replay methods are the most effective method.

Simple and effective deep hand shape and pose regression from a single depth image

We revisit Semantic Scene Completion (SSC), a useful task to predict the semantic and occupancy representation of 3D scenes, in this paper. A number of methods for this task are always based on voxelized scene representations. Although voxel representations keep local structures of the scene, these methods suffer from heavy computation redundancy due to the existence of visible empty voxels when the network goes deeper. To address this dilemma, we propose our novel point-voxel aggregation network for this task. We first transfer the voxelized scenes to point clouds by removing these visible empty voxels and adopt a deep point stream to capture semantic information from the scene efficiently. Meanwhile, a light-weight voxel stream containing only two 3D convolution layers preserves local structures of the voxelized scenes. Furthermore, we design an anisotropic voxel aggregation operator to fuse the structure details from the voxel stream into the point stream, and a semantic-aware propagation module to enhance the up-sampling process in the point stream by semantic labels. We demonstrate that our model surpasses state-of-the-arts on two benchmarks by a large margin, with only the depth images as input.

Face recognition has been one of the most studied researches in computer vision, and facial feature extraction, is one of the cores of face recognition. In this paper, we focus on semantic segmentation methods for facial feature extraction. However, the structure and number of parameters of the model usually affect the accuracy of semantic segmentation tasks in facial feature extraction. High-resolution representation learning can bring better accuracy to semantic segmentation, and large convolutional kernels, which bring a larger effective receptive field and better shape feature extraction to the model, can also improve the performance of the segmentation model. In this paper, we propose a novel lightweight model, HRModel (HRM), for optimizing the accuracy and parameters of semantic segmentation tasks in face recognition. From the perspective of multi-scale fusion, large perceptual fields, and gradient segmentation using high-resolution parallel streaming networks as a framework, we use a multi-scale feature extraction module with a large convolutional kernel to reduce the number of parameters while maintaining accuracy. In the feature extraction block, we use a channel attention mechanism to aggregate feature maps of different depths and different receptive fields. At the same time, we apply depthwise separable convolution and gradient flow splitting to reuse the gradient information and hence reduce the number of parameters and computational effort of the network effectively. We compare our proposed model HRM with other single-model networks on semantic segmentation tasks for face recognition. The experimental results show that HRM significantly reduces the number of parameters and achieves an accuracy of 95.52% in the application of the CelebAMask-HQ dataset.

While current approaches for neural network training often aim at improving performance, less focus is put on training methods aiming at robustness towards varying noise conditions or directed attacks by adversarial examples. In this paper, we propose to improve robustness by a multi-task training, which extends supervised semantic segmentation by a self-supervised monocular depth estimation on unlabeled videos. This additional task is only performed during training to improve the semantic segmentation model's robustness at test time under several input perturbations. Moreover, we even find that our joint training approach also improves the performance of the model on the original (supervised) semantic segmentation task. Our evaluation exhibits a particular novelty in that it allows to mutually compare the effect of input noises and adversarial attacks on the robustness of the semantic segmentation. We show the effectiveness of our method on the Cityscapes dataset, where our multi-task training approach consistently outperforms the single-task semantic segmentation baseline in terms of both robustness vs. noise and in terms of adversarial attacks, without the need for depth labels in training.

Depth images are widely used in 3D head pose estimation and face reconstruction. The device-specific noise and the lack of textual constraints pose a major problem for estimating a nonrigid deformable face from a single noisy depth image. In this article, we present a deep neural network-based framework to infer a 3D face consistent with a single depth image captured by a consumer depth camera Kinect. Confronted with a lack of annotated depth images with facial parameters, we utilize the bidirectional CycleGAN-based generator for denoising and noisy image simulation, which helps to generalize the model learned from synthetic depth images to real noisy ones. We generate the code regressors in the source (synthetic) and the target (noisy) depth image domains and present a fusion scheme in the parametric space for 3D face inference. The proposed multi-level shape consistency constraint, concerning the embedded features, depth maps, and 3D surfaces, couples the code regressor and the domain adaptation, avoiding shape distortions in the CycleGAN-based generators. Experiments demonstrate that the proposed method is effective in depth-based 3D head pose estimation and expressive face reconstruction compared with the state-of-the-art.