Challenging the Universal Representation of Deep Models for 3D Point Cloud Registration

Learning universal representations across different applications domain is an open research problem. In fact, finding universal architecture within the same application but across different types of datasets is still unsolved problem too, especially in applications involving processing 3D point clouds. In this work we experimentally test several state-of-the-art learning-based methods for 3D point cloud registration against the proposed non-learning baseline registration method. The proposed method either outperforms or achieves comparable results w.r.t. learning based methods. In addition, we propose a dataset on which learning based methods have a hard time to generalize. Our proposed method and dataset, along with the provided experiments, can be used in further research in studying effective solutions for universal representations. Our source code is available at: github.com/DavidBoja/greedy-grid-search.

Stochastic Modeling for Learnable Human Pose Triangulation

We propose a stochastic modeling framework for 3D human pose triangulation and evaluate its performance across different datasets and spatial camera arrangements. The common approach to 3D pose estimation is to first detect 2D keypoints in images and then apply the triangulation from multiple views. However, the majority of existing triangulation models are limited to a single dataset, i.e. camera arrangement and their number. Moreover, they require known camera parameters. The proposed stochastic pose triangulation model successfully generalizes to different camera arrangements and between two public datasets. In each step, we generate a set of 3D pose hypotheses obtained by triangulation from a random subset of views. The hypotheses are evaluated by a neural network and the expectation of the triangulation error is minimized. The key novelty is that the network learns to evaluate the poses without taking into account the spatial camera arrangement, thus improving generalization. Additionally, we demonstrate that the proposed stochastic framework can also be used for fundamental matrix estimation, showing promising results towards relative camera pose estimation from noisy keypoint correspondences.

Catadioptric Stereo on a Smartphone

We present a 3D printed adapter with planar mirrors for stereo reconstruction using front and back smartphone camera. The adapter presents a practical and low-cost solution for enabling any smartphone to be used as a stereo camera, which is currently only possible using high-end phones with expensive 3D sensors. Using the prototype version of the adapter, we experiment with parameters like the angles between cameras and mirrors and the distance to each camera (the stereo baseline). We find the most convenient configuration and calibrate the stereo pair. Based on the presented preliminary analysis, we identify possible improvements in the current design. To demonstrate the working prototype, we reconstruct a 3D human pose using 2D keypoint detections from the stereo pair and evaluate extracted body lengths. The result shows that the adapter can be used for anthropometric measurement of several body segments.

Can Human Sex Be Learned Using Only 2D Body Keypoint Estimations?

In this paper, we analyze human male and female sex recognition problem and present a fully automated classification system using only 2D keypoints. The keypoints represent human joints. A keypoint set consists of 15 joints and the keypoint estimations are obtained using an OpenPose 2D keypoint detector. We learn a deep learning model to distinguish males and females using the keypoints as input and binary labels as output. We use two public datasets in the experimental section - 3DPeople and PETA. On PETA dataset, we report a 77% accuracy. We provide model performance details on both PETA and 3DPeople. To measure the effect of noisy 2D keypoint detections on the performance, we run separate experiments on 3DPeople ground truth and noisy keypoint data. Finally, we extract a set of factors that affect the classification accuracy and propose future work. The advantage of the approach is that the input is small and the architecture is simple, which enables us to run many experiments and keep the real-time performance in inference. The source code, with the experiments and data preparation scripts, are available on GitHub (https://github.com/kristijanbartol/human-sex-classifier).

Smart Time-Multiplexing of Quads Solves the Multicamera Interference Problem

Time-of-flight (ToF) cameras are becoming increasingly popular for 3D imaging. Their optimal usage has been studied from the several aspects. One of the open research problems is the possibility of a multicamera interference problem when two or more ToF cameras are operating simultaneously. In this work we present an efficient method to synchronize multiple operating ToF cameras. Our method is based on the time-division multiplexing, but unlike traditional time multiplexing, it does not decrease the effective camera frame rate. Additionally, for unsynchronized cameras, we provide a robust method to extract from their corresponding video streams, frames which are not subject to multicamera interference problem. We demonstrate our approach through a series of experiments and with a different level of support available for triggering, ranging from a hardware triggering to purely random software triggering.

Towards Keypoint Guided Self-Supervised Depth Estimation

This paper proposes to use keypoints as a self-supervision clue for learning depth map estimation from a collection of input images. As ground truth depth from real images is difficult to obtain, there are many unsupervised and self-supervised approaches to depth estimation that have been proposed. Most of these unsupervised approaches use depth map and ego-motion estimations to reproject the pixels from the current image into the adjacent image from the image collection. Depth and ego-motion estimations are evaluated based on pixel intensity differences between the correspondent original and reprojected pixels. Instead of reprojecting the individual pixels, we propose to first select image keypoints in both images and then reproject and compare the correspondent keypoints of the two images. The keypoints should describe the distinctive image features well. By learning a deep model with and without the keypoint extraction technique, we show that using the keypoints improve the depth estimation learning. We also propose some future directions for keypoint-guided learning of structure-from-motion problems.

On the Comparison of Classic and Deep Keypoint Detector and Descriptor Methods

The purpose of this study is to give a performance comparison between several classic hand-crafted and deep key-point detector and descriptor methods. In particular, we consider the following classical algorithms: SIFT, SURF, ORB, FAST, BRISK, MSER, HARRIS, KAZE, AKAZE, AGAST, GFTT, FREAK, BRIEF and RootSIFT, where a subset of all combinations is paired into detector-descriptor pipelines. Additionally, we analyze the performance of two recent and perspective deep detector-descriptor models, LF-Net and SuperPoint. Our benchmark relies on the HPSequences dataset that provides real and diverse images under various geometric and illumination changes. We analyze the performance on three evaluation tasks: keypoint verification, image matching and keypoint retrieval. The results show that certain classic and deep approaches are still comparable, with some classic detector-descriptor combinations overperforming pretrained deep models. In terms of the execution times of tested implementations, SuperPoint model is the fastest, followed by ORB. The source code is published on \url{https://github.com/kristijanbartol/keypoint-algorithms-benchmark}.