Current learning-based Structure-from-Motion (SfM) methods struggle with videos of dynamic scenes captured by wide-angle cameras. We present anyMap, a differentiable SfM framework that jointly addresses image distortion and motion estimation. By learning a general implicit camera model without predefined parameters, anyMap handles lens distortion and estimates multi-view consistent 3D geometry, camera poses, and (un)projection functions. To resolve the ambiguity where motion estimation can compensate for undistortion errors and vice versa, we introduce a low-dimensional motion representation consisting of a set of learnable basis trajectories, which are interpolated to produce regularized motion estimates. Experimental results show that our method achieves accurate camera poses, excels in camera calibration and image rectification, and enables high-quality novel view synthesis. Our low-dimensional motion representation effectively disentangles undistortion from motion estimation, outperforming existing methods.

A key challenge in neural 3D scene reconstruction from monocular images is to fuse features back projected from various views without any depth or occlusion information. We address this by leveraging monocular depth priors, which effectively guide the fusion to improve surface prediction and skip over irrelevant, ambiguous, or occluded features. Furthermore, we revisit the average-based fusion used by most neural 3D reconstruction methods and propose two alternatives, a variance-based and a cross-attention-based fusion module, that are more efficient and effective than the average-based and self-attention-based counterparts. Compared to the NeuralRecon baseline, the proposed DG-Recon models significantly improve the reconstruction quality and completeness while remaining in real-time. Our method achieves state-of-the-art online reconstruction results on the ScanNet dataset and is on par with the current best offline method, which repeatedly accesses keyframes from the entire video sequence. Our ScanNet-trained model also generalizes robustly to the challenging 7-Scenes dataset and a subset of SUN3D containing scenes as big as an entire floor.

Self-supervised monocular depth estimation (SSMDE) aims at predicting the dense depth maps of monocular images, by learning to minimize a photometric loss using spatially neighboring image pairs during training. While SSMDE offers a significant scalability advantage over supervised approaches, it performs poorly on reflective surfaces as the photometric constancy assumption of the photometric loss is violated. We note that the appearance of reflective surfaces is view-dependent and often there are views of such surfaces in the training data that are not contaminated by strong specular reflections. Thus, reflective surfaces can be accurately reconstructed by aggregating the predicted depth of these views. Motivated by this observation, we propose 3D distillation: a novel training framework that utilizes the projected depth of reconstructed reflective surfaces to generate reasonably accurate depth pseudo-labels. To identify those surfaces automatically, we employ an uncertainty-guided depth fusion method, combining the smoother and more accurate projected depth on reflective surfaces and the detailed predicted depth elsewhere. In our experiments using the ScanNet and 7-Scenes datasets, we show that 3D distillation not only significantly improves the prediction accuracy, especially on the problematic surfaces, but also that it generalizes well over various underlying network architectures and to new datasets.

In semantic segmentation tasks, input images can often have more than one plausible interpretation, thus allowing for multiple valid labels. To capture such ambiguities, recent work has explored the use of probabilistic networks that can learn a distribution over predictions. However, these do not necessarily represent the empirical distribution accurately. In this work, we present a strategy for learning a calibrated predictive distribution over semantic maps, where the probability associated with each prediction reflects its ground truth correctness likelihood. To this end, we propose a novel two-stage, cascaded approach for calibrated adversarial refinement: (i) a standard segmentation network is trained with categorical cross entropy to predict a pixelwise probability distribution over semantic classes and (ii) an adversarially trained stochastic network is used to model the inter-pixel correlations to refine the output of the first network into coherent samples. Importantly, to calibrate the refinement network and prevent mode collapse, the expectation of the samples in the second stage is matched to the probabilities predicted in the first. We demonstrate the versatility and robustness of the approach by achieving state-of-the-art results on the multigrader LIDC dataset and on a modified Cityscapes dataset with injected ambiguities. In addition, we show that the core design can be adapted to other tasks requiring learning a calibrated predictive distribution by experimenting on a toy regression dataset. We provide an open source implementation of our method at https://github.com/EliasKassapis/CARSSS.

Predicting the biological function of molecules, be it proteins or drug-like compounds, from their atomic structure is an important and long-standing problem. The electron density field and electrostatic potential field of a molecule contain the “raw fingerprint” of how this molecule can fit to binding partners. In this paper, we show that deep learning can predict biological function of molecules directly from their raw 3D approximated electron density and electrostatic potential fields. Protein function based on Enzyme Commission numbers is predicted from the approximated electron density field. In another experiment, the activity of small molecules is predicted with quality comparable to state-of-the-art descriptor-based methods. We propose several alternative computational models for the GPU with different memory and runtime requirements for different sizes of molecules and of databases. We also propose application-specific multi-channel data representations.

In this paper we propose a Bayesian method for estimating architectural parameters of neural networks, namely layer size and network depth. We do this by learning concrete distributions over these parameters. Our results show that regular networks with a learned structure can generalise better on small datasets, while fully stochastic networks can be more robust to parameter initialisation. The proposed method relies on standard neural variational learning and, unlike randomised architecture search, does not require a retraining of the model, thus keeping the computational overhead at minimum.