Xue-qian Li

Neural Scene Flow Prior (NSFP) is of significant interest to the vision community due to its inherent robustness to out-of-distribution (OOD) effects and its ability to deal with dense lidar points. The approach utilizes a coordinate neural network to estimate scene flow at runtime, without any training. However, it is up to 100 times slower than current state-of-the-art learning methods. We rediscover the distance transform (DT) as an efficient, correspondence-free loss function that dramatically speeds up the runtime optimization, allowing for the first time real-time performance comparable to learning methods.

We use non-Fourier positional encodings (e.g., shifted Gaussian functions) to show the signal reconstruction is determined by a trade-off between the stable rank of the embedding and the distance preservation between embedded coordinates. Furthermore, employing a complex positional encoding requires only a linear (rather than deep) function to achieve comparable performance, while being orders of magnitude faster than current state-of-the-art.

Traditionally, trajectories have been represented by a set of handcrafted bases that have limited expressibility. Here, we propose a neural trajectory prior to capture continuous spatio-temporal information without the need for offline data. We demonstrate how our proposed objective is optimized during runtime to estimate trajectories for two important tasks: Non-Rigid Structure from Motion (NRSfM) and lidar scene flow integration for self-driving scenes.

We propose a neural scene flow prior that utilizes the architecture of neural networks (MLPs) as an implicit regularizer. Unlike learning-based scene flow methods, optimization occurs at runtime, and our approach needs no offline datasets -- making it ideal for deployment in new environments such as autonomous driving. Also, the implicit and continuous scene flow representation allows us to estimate dense long-term correspondences across a sequence of point clouds.

We revisit a recent innovation -- PointNetLK -- and show that the inclusion of an analytical Jacobian can exhibit remarkable generalization properties while reaping the inherent fidelity benefits of a learning framework. Our approach not only outperforms the state-of-the-art in mismatched conditions but also produces results competitive with current learning methods when operating on real-world test data close to the training set.

We provide a novel network to align point clouds that utilizes PointNet encoding. Our PCRNet uses data specific information to make accurate registration estimation, which is robust to noise and large initial misalignment between source point cloud and template point cloud. With single pass, PCRNet performs computationally faster than existing methods, while iterative version of PCRNet can refine the registration result to make highly accurate prediction. Our PCRNet can be applied further to 3D tracking, model replacement.

We propose a multi-domain feature learning-based visual place recognition method with 2 modules: feature extraction module, which uses CapsuleNet to extract features under different circumstances; feature separation module, which enforces a separation in environmental condition-invariant and condition-related domains. Combining 2 modules through sequential matching, we can find potential match.

In order to perform sequence matching between short-term testing frame and long-term reference frame in visual place recognition, we develop a multi-resolution sampling-based global search method with coarse-to-fine searching and particle filter, which balances the matching accuracy and searching efficiency. By iteratively updating both potential trajectories and frame sequences through low resolution to high resolution, we can find the best match under the highest resolution level.