Ryan Farrell

Connor Anderson, Ryan Farrell

The deep neural networks used in modern computer vision systems require enormous image datasets to train them. These carefully-curated datasets typically have a million or more images, across a thousand or more distinct categories. The process of creating and curating such a dataset is a monumental undertaking, demanding extensive effort and labelling expense and necessitating careful navigation of technical and social issues such as label accuracy, copyright ownership, and content bias. What if we had a way to harness the power of large image datasets but with few or none of the major issues and concerns currently faced? This paper extends the recent work of Kataoka et. al. (2020), proposing an improved pre-training dataset based on dynamically-generated fractal images. Challenging issues with large-scale image datasets become points of elegance for fractal pre-training: perfect label accuracy at zero cost; no need to store/transmit large image archives; no privacy/demographic bias/concerns of inappropriate content, as no humans are pictured; limitless supply and diversity of images; and the images are free/open-source. Perhaps surprisingly, avoiding these difficulties imposes only a small penalty in performance. Leveraging a newly-proposed pre-training task -- multi-instance prediction -- our experiments demonstrate that fine-tuning a network pre-trained using fractals attains 92.7-98.1% of the accuracy of an ImageNet pre-trained network.

Pei Guo, Ryan Farrell

In this work, we explore the task of semantic object keypoint discovery weakly-supervised by only category labels. This is achieved by transforming discriminatively-trained intermediate layer filters into keypoint detectors. We begin by identifying three preferred characteristics of keypoint detectors: (i) spatially sparse activations, (ii) consistency and (iii) diversity. Instead of relying on hand-crafted loss terms, a novel computationally-efficient leaky max pooling (LMP) layer is proposed to explicitly encourage final conv-layer filters to learn "non-repeatable local patterns" that are well aligned with object keypoints. Informed by visualizations, a simple yet effective selection strategy is proposed to ensure consistent filter activations and attention mask-out is then applied to force the network to distribute its attention to the whole object instead of just the most discriminative region. For the final keypoint prediction, a learnable clustering layer is proposed to group keypoint proposals into keypoint predictions. The final model, named LMPNet, is highly interpretable in that it directly manipulates network filters to detect predefined concepts. Our experiments show that LMPNet can (i) automatically discover semantic keypoints that are robust to object pose and (ii) achieves strong prediction accuracy comparable to a supervised pose estimation model.

Matthew Gwilliam, Adam Teuscher, Connor Anderson et al.

For the task of image classification, researchers work arduously to develop the next state-of-the-art (SOTA) model, each bench-marking their own performance against that of their predecessors and of their peers. Unfortunately, the metric used most frequently to describe a model's performance, average categorization accuracy, is often used in isolation. As the number of classes increases, such as in fine-grained visual categorization (FGVC), the amount of information conveyed by average accuracy alone dwindles. While its most glaring weakness is its failure to describe the model's performance on a class-by-class basis, average accuracy also fails to describe how performance may vary from one trained model of the same architecture, on the same dataset, to another (both averaged across all categories and at the per-class level). We first demonstrate the magnitude of these variations across models and across class distributions based on attributes of the data, comparing results on different visual domains and different per-class image distributions, including long-tailed distributions and few-shot subsets. We then analyze the impact various FGVC methods have on overall and per-class variance. From this analysis, we both highlight the importance of reporting and comparing methods based on information beyond overall accuracy, as well as point out techniques that mitigate variance in FGVC results.

Connor Anderson, Matt Gwilliam, Adam Teuscher et al.

In fine-grained visual categorization (FGVC), there is a near-singular focus in pursuit of attaining state-of-the-art (SOTA) accuracy. This work carefully analyzes the performance of recent SOTA methods, quantitatively, but more importantly, qualitatively. We show that these models universally struggle with certain "hard" images, while also making complementary mistakes. We underscore the importance of such analysis, and demonstrate that combining complementary models can improve accuracy on the popular CUB-200 dataset by over 5%. In addition to detailed analysis and characterization of the errors made by these SOTA methods, we provide a clear set of recommended directions for future FGVC researchers.

Pei Guo, Ryan Farrell

Network interpretation as an effort to reveal the features learned by a network remains largely visualization-based. In this paper, our goal is to tackle semantic network interpretation at both filter and decision level. For filter-level interpretation, we represent the concepts a filter encodes with a probability distribution of visual attributes. The decision-level interpretation is achieved by textual summarization that generates an explanatory sentence containing clues behind a network's decision. A Bayesian inference algorithm is proposed to automatically associate filters and network decisions with visual attributes. Human study confirms that the semantic interpretation is a beneficial alternative or complement to visualization methods. We demonstrate the crucial role that semantic network interpretation can play in understanding a network's failure patterns. More importantly, semantic network interpretation enables a better understanding of the correlation between a model's performance and its distribution metrics like filter selectivity and concept sparseness.

Pei Guo, Ryan Farrell

Dramatic appearance variation due to pose constitutes a great challenge in fine-grained recognition, one which recent methods using attention mechanisms or second-order statistics fail to adequately address. Modern CNNs typically lack an explicit understanding of object pose and are instead confused by entangled pose and appearance. In this paper, we propose a unified object representation built from a hierarchy of pose-aligned regions. Rather than representing an object by regions aligned to image axes, the proposed representation characterizes appearance relative to the object's pose using pose-aligned patches whose features are robust to variations in pose, scale and rotation. We propose an algorithm that performs pose estimation and forms the unified object representation as the concatenation of hierarchical pose-aligned regions features, which is then fed into a classification network. The proposed algorithm surpasses the performance of other approaches, increasing the state-of-the-art by nearly 2% on the widely-used CUB-200 dataset and by more than 8% on the much larger NABirds dataset. The effectiveness of this paradigm relative to competing methods suggests the critical importance of disentangling pose and appearance for continued progress in fine-grained recognition.

Abhimanyu Dubey, Otkrist Gupta, Pei Guo et al.

Fine-Grained Visual Classification (FGVC) datasets contain small sample sizes, along with significant intra-class variation and inter-class similarity. While prior work has addressed intra-class variation using localization and segmentation techniques, inter-class similarity may also affect feature learning and reduce classification performance. In this work, we address this problem using a novel optimization procedure for the end-to-end neural network training on FGVC tasks. Our procedure, called Pairwise Confusion (PC) reduces overfitting by intentionally {introducing confusion} in the activations. With PC regularization, we obtain state-of-the-art performance on six of the most widely-used FGVC datasets and demonstrate improved localization ability. {PC} is easy to implement, does not need excessive hyperparameter tuning during training, and does not add significant overhead during test time.