Towards Generalized Frameworks for Object Recognition

Abstract

Over the past few years, deep convolutional neural network (DCNN) based approaches have been immensely successful in tackling a diverse range of object recognition problems. Popular DCNN architectures like deep residual networks (ResNets) are highly generic, not just for classification, but also for high level tasks like detection/tracking which rely on classification DCNNs as their backbone. The generality of DCNNs however doesn't extend to image-to-image(Im2Im) regression tasks (eg: super-resolution, denoising, rgb-to-depth, relighting, etc). For such tasks, DCNNs are often highly task-specific and require specific ancillary post-processing methods. The major issue plaguing the design of generic architectures for such tasks is the tradeoff between context/locality given a fixed computation/memory budget. We first present a generic DCNN architecture for Im2Im regression that can be trained end-to-end without any further machinery. Our proposed architecture, the Recursively Branched Deconvolutional Network (RBDN), which features a cheap early multi-context image representation, an efficient recursive branching scheme with extensive parameter sharing and learnable upsampling. We provide qualitative/quantitative results on 3 diverse tasks: relighting, denoising and colorization and show that our proposed RBDN architecture obtains comparable results to the state-of-the-art on each of these tasks when used off-the-shelf without any post processing or task-specific architectural modifications. Second, we focus on gradient flow and optimization in ResNets. In particular, we theoretically analyze why pre-activation(v2) ResNets outperform the original ResNets(v1) on CIFAR datasets but not on ImageNet. Our analysis reveals that although v1-ResNets lack ensembling properties, they can have a higher effective depth in comparison to v2-ResNes. Subsequently, we show that downsampling projections (while only few in number) have a significantly detrimental effect on performance. We show that by simply replacing downsampling-projections with identity-like dense-reshape shortcuts, the classification results of standard residual architectures like ResNets, ResNeXts and SE-Nets improve by up to 1.2% on ImageNet, without any increase in computational complexity (FLOPs). Finally, we present a robust non-parametric probabilistic ensemble method for multi-classification, which outperforms the state-of-the-art ensemble methods on several machine learning and computer vision datasets for object recognition with statistically significant improvements. The approach is particularly geared towards multi-classification problems with very low training data and/or a fairly high proportion of outliers, for which training end-to-end DCNNs is not very beneficial.