CAT: Contrastive Adapter Training for Personalized Image Generation

The emergence of various adapters, including Low-Rank Adaptation (LoRA) applied from the field of natural language processing, has allowed diffusion models to personalize image generation at a low cost. However, due to the various challenges including limited datasets and shortage of regularization and computation resources, adapter training often results in unsatisfactory outcomes, leading to the corruption of the backbone model's prior knowledge. One of the well known phenomena is the loss of diversity in object generation, especially within the same class which leads to generating almost identical objects with minor variations. This poses challenges in generation capabilities. To solve this issue, we present Contrastive Adapter Training (CAT), a simple yet effective strategy to enhance adapter training through the application of CAT loss. Our approach facilitates the preservation of the base model's original knowledge when the model initiates adapters. Furthermore, we introduce the Knowledge Preservation Score (KPS) to evaluate CAT's ability to keep the former information. We qualitatively and quantitatively compare CAT's improvement. Finally, we mention the possibility of CAT in the aspects of multi-concept adapter and optimization.

Self-supervised Pre-training for Precipitation Post-processor

Securing sufficient forecast lead time for local precipitation is essential for preventing hazardous weather events. Nonetheless, global warming-induced climate change is adding to the challenge of accurately predicting severe precipitation events, such as heavy rainfall. In this work, we propose a deep learning-based precipitation post-processor approach to numerical weather prediction (NWP) models. The precipitation post-processor consists of (i) self-supervised pre-training, where parameters of encoder are pre-trained on the reconstruction of masked variables of the atmospheric physics domain, and (ii) transfer learning on precipitation segmentation tasks (target domain) from the pre-trained encoder. We also introduce a heuristic labeling approach for effectively training class-imbalanced datasets. Our experiment results in precipitation correction for regional NWP show that the proposed method outperforms other approaches.

PeRFception: Perception using Radiance Fields

The recent progress in implicit 3D representation, i.e., Neural Radiance Fields (NeRFs), has made accurate and photorealistic 3D reconstruction possible in a differentiable manner. This new representation can effectively convey the information of hundreds of high-resolution images in one compact format and allows photorealistic synthesis of novel views. In this work, using the variant of NeRF called Plenoxels, we create the first large-scale implicit representation datasets for perception tasks, called the PeRFception, which consists of two parts that incorporate both object-centric and scene-centric scans for classification and segmentation. It shows a significant memory compression rate (96.4\%) from the original dataset, while containing both 2D and 3D information in a unified form. We construct the classification and segmentation models that directly take as input this implicit format and also propose a novel augmentation technique to avoid overfitting on backgrounds of images. The code and data are publicly available in https://postech-cvlab.github.io/PeRFception .

Learning to Register Unbalanced Point Pairs

Recent 3D registration methods can effectively handle large-scale or partially overlapping point pairs. However, despite its practicality, matching the unbalanced pairs in terms of spatial scale and density has been overlooked. We present a novel 3D registration method, called UPPNet, for the unbalanced point pairs. We propose a hierarchical framework to find inlier correspondences effectively by gradually reducing search space. Our method predicts the subregions of the target points likely to be overlapped with the query points. The following super-point matching module and fine-grained refinement module estimate accurate inlier correspondences between two point clouds. Furthermore, we apply geometric constraints to refine the correspondences that satisfy spatial compatibility. Correspondence prediction is trained end-to-end, and our approach can predict the proper rigid transformation with a single forward pass given unbalanced point cloud pairs. To validate the efficacy of the proposed method, we create a KITTI-UPP dataset by augmenting the KITTI LiDAR dataset. Experiments on this dataset reveal that the proposed approach significantly outperforms state-of-the-art pairwise point cloud registration methods by a large margin, resulting in 78% improvement in Registration Recall when the target point cloud is about 10$\times$ spatially larger and about 10$\times$ times denser than the query point cloud.

Putting 3D Spatially Sparse Networks on a Diet

3D neural networks have become prevalent for many 3D vision tasks including object detection, segmentation, registration, and various perception tasks for 3D inputs. However, due to the sparsity and irregularity of 3D data, custom 3D operators or network designs have been the primary focus of 3D research, while the size of networks or efficacy of parameters has been overlooked. In this work, we perform the first comprehensive study on the weight sparsity of spatially sparse 3D convolutional networks and propose a compact weight-sparse and spatially sparse 3D convnet (WS^3-ConvNet) for semantic segmentation and instance segmentation. We employ various network pruning strategies to find compact networks and show our WS^3-ConvNet achieves minimal loss in performance (2.15% drop) with orders-of-magnitude smaller number of parameters (1/100 compression rate). Finally, we systematically analyze the compression patterns of WS^3-ConvNet and show interesting emerging sparsity patterns common in our compressed networks to further speed up inference.

Deep Hough Voting for Robust Global Registration

Point cloud registration is the task of estimating the rigid transformation that aligns a pair of point cloud fragments. We present an efficient and robust framework for pairwise registration of real-world 3D scans, leveraging Hough voting in the 6D transformation parameter space. First, deep geometric features are extracted from a point cloud pair to compute putative correspondences. We then construct a set of triplets of correspondences to cast votes on the 6D Hough space, representing the transformation parameters in sparse tensors. Next, a fully convolutional refinement module is applied to refine the noisy votes. Finally, we identify the consensus among the correspondences from the Hough space, which we use to predict our final transformation parameters. Our method outperforms state-of-the-art methods on 3DMatch and 3DLoMatch benchmarks while achieving comparable performance on KITTI odometry dataset. We further demonstrate the generalizability of our approach by setting a new state-of-the-art on ICL-NUIM dataset, where we integrate our module into a multi-way registration pipeline.

High-dimensional Convolutional Networks for Geometric Pattern Recognition

Many problems in science and engineering can be formulated in terms of geometric patterns in high-dimensional spaces. We present high-dimensional convolutional networks (ConvNets) for pattern recognition problems that arise in the context of geometric registration. We first study the effectiveness of convolutional networks in detecting linear subspaces in high-dimensional spaces with up to 32 dimensions: much higher dimensionality than prior applications of ConvNets. We then apply high-dimensional ConvNets to 3D registration under rigid motions and image correspondence estimation. Experiments indicate that our high-dimensional ConvNets outperform prior approaches that relied on deep networks based on global pooling operators.