Eyeglasses removal is challenging in removing different kinds of eyeglasses, e.g., rimless glasses, full-rim glasses, and sunglasses, and recovering appropriate eyes. Due to the significant visual variants, the conventional methods lack scalability. Most existing works focus on the frontal face images in the controlled environment, such as the laboratory, and need to design specific systems for different eyeglass types. To address the limitation, we propose a unified eyeglass removal model called the eyeglasses removal generative adversarial network (ERGAN), which could handle different types of glasses in the wild. The proposed method does not depend on the dense annotation of eyeglasses location but benefits from the large-scale face images with weak annotations. Specifically, we study the two relevant tasks simultaneously, that is, removing eyeglasses and wearing eyeglasses. Given two face images with and without eyeglasses, the proposed model learns to swap the eye area in two faces. The generation mechanism focuses on the eye area and invades the difficulty of generating a new face. In the experiment, we show the proposed method achieves a competitive removal quality in terms of realism and diversity. Furthermore, we evaluate ERGAN on several subsequent tasks, such as face verification and facial expression recognition. The experiment shows that our method could serve as a preprocessing method for these tasks.

In the point-cloud-based place recognition area, the existing hybrid architectures combining both convolutional networks and transformers have shown promising performance. They mainly apply the voxel-wise transformer after the sparse convolution (SPConv). However, they can induce information loss by the sparse voxelization and further result in loss propagation to the transformer, significantly degrading the performance of the network, especially in outdoor scenes with complex geometric structures and multiple small objects. To address this issue, we propose a novel Point-wise Transformer with sparse Convolution (PTC). Specifically, SPConv is applied to the sparsely voxelized point cloud to extract local features, which are then converted to the point-based representation via a feature transformation unit (FTU). As such, our PTC can apply a transformer model based on the point-wise representation rather than on the voxel-wise one. To enhance the ability to capture long-range features and reduce the computational complexity of the transformer, we propose a two-step transformer, each with different grouping strategies. Meanwhile, in both steps, the attention matrix is computed with much fewer points by grouping a single point cloud into different attention domains. The experiments show that the PTC-Net can achieve state-of-the-art (SOTA) performance, with an improvement of 3.6% on average recall@1. Furthermore, to demonstrate the effectiveness of the PTC, we introduce an extremely light-weight version, PTC-Net-L, with only one PTC layer and half initial channel dimensions, also achieving SOTA performance in terms of the average recall rate and running time with only 0.08 M parameters.

Road Detection is a basic task in automated driving field, in which 3D lidar data is commonly used recently. In this paper, we propose to rearrange 3D lidar data into a new organized form to construct direct spatial relationship among point cloud, and put forward new features for real-time road detection tasks. Our model works based on two prerequisites: (1) Road regions are always flatter than non-road regions. (2) Light travels in straight lines in a uniform medium. Based on prerequisite 1, we put forward difference-between-lines feature, while ScanID density and obstacle radial map are generated based on prerequisite 2. According to our method, we construct an array of structures to store and reorganize 3D input firstly. Then, two novel features, difference-between-lines and ScanID density, are extracted, based on which we construct a consistency map and an obstacle map in Bird Eye View (BEV). Finally, the road region is extracted by fusing these two maps and refinement is used to polish up our outcome. We have carried out experiments on the public KITTI-Road benchmark, achieving one of the best performances among the lidar-based road detection methods. To further prove the efficiency of our method on unstructured road, the visual outcomes in rural areas are also proposed.

Multi-object tracking aims to recover object trajectories given multiple detections in video frames. Object feature extraction and similarity metric are two keys to reliably associate trajectories. In this paper, we propose the recurrent metric network (RMNet), a CNN-RNN based similarity metric framework for multi-object tracking. Given a reference object, the RMNet takes as input random positive and negative detections and outputs similarity scores over time. The RMNet handles the long-term temporal object variations and false object detections by its long-short memory units. With the scores from RMNet, we introduce a batch multiple hypothesis (BMH) strategy, a simple yet efficient data association method for batch multi-object tracking. BMH generates a hypothesis tree for each object with a dual-threshold hypothesis generation approach, and then selects the best branch (or hypothesis) for each object as the batch tracking result. Specially, we model hypothesis selection as a 0–1 programming problem and introduce a reward function to re-find objects in case of missing detection.We evaluate our RMNet and BMH strategy on several popular datasets: 2DMOT2015, PETS2009, TUD, and KITTI. We achieve performance comparable or superior to these of the state-of-the-art methods.