We tackle the challenge of learning a distribution over complex, realistic, indoor scenes. In this paper, we introduce Generative Scene Networks (GSN), which learns to decompose scenes into a collection of many local radiance fields that can be rendered from a free moving camera. Our model can be used as a prior to generate new scenes, or to complete a scene given only sparse 2D observations. Recent work has shown that generative models of radiance fields can capture properties such as multi-view consistency and view-dependent lighting. However, these models are specialized for constrained viewing of single objects, such as cars or faces. Due to the size and complexity of realistic indoor environments, existing models lack the representational capacity to adequately capture them. Our decomposition scheme scales to larger and more complex scenes while preserving details and diversity, and the learned prior enables high-quality rendering from viewpoints that are significantly different from observed viewpoints. When compared to existing models, GSN produces quantitatively higher-quality scene renderings across several different scene datasets.
We consider the task of semi-supervised semantic segmentation, where we aim to produce pixel-wise semantic object masks given only a small number of human-labeled training examples. We focus on iterative self-training methods in which we explore the behavior of self-training over multiple refinement stages. We show that iterative self-training leads to performance degradation if done naively with a fixed ratio of human-labeled to pseudo-labeled training examples. We propose Greedy Iterative Self-Training (GIST) and Random Iterative Self-Training (RIST) strategies that alternate between training on either human-labeled data or pseudo-labeled data at each refinement stage, resulting in a performance boost rather than degradation. We further show that GIST and RIST can be combined with existing SOTA methods to boost performance, yielding new SOTA results in Pascal VOC 2012 and Cityscapes dataset across five out of six subsets.
Hairstyle transfer is challenging due to hair structure differences in the source and target hair. Therefore, we propose Latent Optimization of Hairstyles via Orthogonalization (LOHO), an optimization-based approach using GAN inversion to infill missing hair structure details in latent space during hairstyle transfer. Our approach decomposes hair into three attributes: perceptual structure, appearance, and style, and includes tailored losses to model each of these attributes independently. Furthermore, we propose two-stage optimization and gradient orthogonalization to enable disentangled latent space optimization of our hair attributes. Using LOHO for latent space manipulation, users can synthesize novel photorealistic images by manipulating hair attributes either individually or jointly, transferring the desired attributes from reference hairstyles. LOHO achieves a superior FID compared with the current state-of-the-art (SOTA) for hairstyle transfer. Additionally, LOHO preserves the subject's identity comparably well according to PSNR and SSIM when compared to SOTA image embedding pipelines. Code is available at https://github.com/dukebw/LOHO.
In this paper we introduce a Transformer-based approach to video object segmentation (VOS). To address compounding error and scalability issues of prior work, we propose a scalable, end-to-end method for VOS called Sparse Spatiotemporal Transformers (SST). SST extracts per-pixel representations for each object in a video using sparse attention over spatiotemporal features. Our attention-based formulation for VOS allows a model to learn to attend over a history of multiple frames and provides suitable inductive bias for performing correspondence-like computations necessary for solving motion segmentation. We demonstrate the effectiveness of attention-based over recurrent networks in the spatiotemporal domain. Our method achieves competitive results on YouTube-VOS and DAVIS 2017 with improved scalability and robustness to occlusions compared with the state of the art.
Generative models are now used to create a variety of high-quality digital artifacts. Yet their use in designing physical objects has received far less attention. In this paper, we advocate for the construction toy, LEGO, as a platform for developing generative models of sequential assembly. We develop a generative model based on graph-structured neural networks that can learn from human-built structures and produce visually compelling designs. Our code is released at: https://github.com/uoguelph-mlrg/GenerativeLEGO.
Neural combinatorial optimization (NCO) aims at designing problem-independent and efficient neural network-based strategies for solving combinatorial problems. The field recently experienced growth by successfully adapting architectures originally designed for machine translation. Even though the results are promising, a large gap still exists between NCO models and classic deterministic solvers, both in terms of accuracy and efficiency. One of the drawbacks of current approaches is the inefficiency of training on multiple problem sizes. Curriculum learning strategies have been shown helpful in increasing performance in the multi-task setting. In this work, we focus on designing a curriculum learning-based training procedure that can help existing architectures achieve competitive performance on a large range of problem sizes simultaneously. We provide a systematic investigation of several training procedures and use the insights gained to motivate application of a psychologically-inspired approach to improve upon the classic curriculum method.
In neuroscience, a tuning dimension is a stimulus attribute that accounts for much of the activation variance of a group of neurons. These are commonly used to decipher the responses of such groups. While researchers have attempted to manually identify an analogue to these tuning dimensions in deep neural networks, we are unaware of an automatic way to discover them. This work contributes an unsupervised framework for identifying and interpreting "tuning dimensions" in deep networks. Our method correctly identifies the tuning dimensions of a synthetic Gabor filter bank and tuning dimensions of the first two layers of InceptionV1 trained on ImageNet.
Recent advances in Generative Adversarial Networks (GANs) have led to their widespread adoption for the purposes of generating high quality synthetic imagery. While capable of generating photo-realistic images, these models often produce unrealistic samples which fall outside of the data manifold. Several recently proposed techniques attempt to avoid spurious samples, either by rejecting them after generation, or by truncating the model's latent space. While effective, these methods are inefficient, as large portions of model capacity are dedicated towards representing samples that will ultimately go unused. In this work we propose a novel approach to improve sample quality: altering the training dataset via instance selection before model training has taken place. To this end, we embed data points into a perceptual feature space and use a simple density model to remove low density regions from the data manifold. By refining the empirical data distribution before training we redirect model capacity towards high-density regions, which ultimately improves sample fidelity. We evaluate our method by training a Self-Attention GAN on ImageNet at 64x64 resolution, where we outperform the current state-of-the-art models on this task while using 1/2 of the parameters. We also highlight training time savings by training a BigGAN on ImageNet at 128x128 resolution, achieving a 66% increase in Inception Score and a 16% improvement in FID over the baseline model with less than 1/4 the training time.
Inferring objects and their relationships from an image is useful in many applications at the intersection of vision and language. Due to a long tail data distribution, the task is challenging, with the inevitable appearance of zero-shot compositions of objects and relationships at test time. Current models often fail to properly understand a scene in such cases, as during training they only observe a tiny fraction of the distribution corresponding to the most frequent compositions. This motivates us to study whether increasing the diversity of the training distribution, by generating replacement for parts of real scene graphs, can lead to better generalization? We employ generative adversarial networks (GANs) conditioned on scene graphs to generate augmented visual features. To increase their diversity, we propose several strategies to perturb the conditioning. One of them is to use a language model, such as BERT, to synthesize plausible yet still unlikely scene graphs. By evaluating our model on Visual Genome, we obtain both positive and negative results. This prompts us to make several observations that can potentially lead to further improvements.
Continual learning is the problem of sequentially learning new tasks or knowledge while protecting previously acquired knowledge. However, catastrophic forgetting poses a grand challenge for neural networks performing such learning process. Thus, neural networks that are deployed in the real world often struggle in scenarios where the data distribution is non-stationary (concept drift), imbalanced, or not always fully available, i.e., rare edge cases. We propose a Differentiable Hebbian Consolidation model which is composed of a Differentiable Hebbian Plasticity (DHP) Softmax layer that adds a rapid learning plastic component (compressed episodic memory) to the fixed (slow changing) parameters of the softmax output layer; enabling learned representations to be retained for a longer timescale. We demonstrate the flexibility of our method by integrating well-known task-specific synaptic consolidation methods to penalize changes in the slow weights that are important for each target task. We evaluate our approach on the Permuted MNIST, Split MNIST and Vision Datasets Mixture benchmarks, and introduce an imbalanced variant of Permuted MNIST -- a dataset that combines the challenges of class imbalance and concept drift. Our proposed model requires no additional hyperparameters and outperforms comparable baselines by reducing forgetting.