Scaling up autoregressive models in vision has not proven as beneficial as in large language models. In this work, we investigate this scaling problem in the context of text-to-image generation, focusing on two critical factors: whether models use discrete or continuous tokens, and whether tokens are generated in a random or fixed raster order using BERT- or GPT-like transformer architectures. Our empirical results show that, while all models scale effectively in terms of validation loss, their evaluation performance -- measured by FID, GenEval score, and visual quality -- follows different trends. Models based on continuous tokens achieves significantly better visual quality than those using discrete tokens. Furthermore, the generation order and attention mechanisms significantly affect the GenEval score: random-order models achieve notably better GenEval scores compared to raster-order models. Inspired by these findings, we train Fluid, a random-order autoregressive model on continuous tokens. Fluid 10.5B model achieves a new state-of-the-art zeor-shot FID of 6.16 on MS-COCO 30K, and 0.69 overall score on the GenEval benchmark. We hope our findings and results will encourage future efforts to further bridge the scaling gap between vision and language models.

We propose Hierarchical Text Spotter (HTS), a novel method for the joint task of word-level text spotting and geometric layout analysis. HTS can recognize text in an image and identify its 4-level hierarchical structure: characters, words, lines, and paragraphs. The proposed HTS is characterized by two novel components: (1) a Unified-DetectorPolygon (UDP) that produces Bezier Curve polygons of text lines and an affinity matrix for paragraph grouping between detected lines; (2) a Line-to-Character-to-Word (L2C2W) recognizer that splits lines into characters and further merges them back into words. HTS achieves stateof-the-art results on multiple word-level text spotting benchmark datasets as well as geometric layout analysis tasks.

Chen-Yu Lee, Chun-Liang Li, Hao Zhang, Timothy Dozat, Vincent Perot, Guolong Su, Xiang Zhang, Kihyuk Sohn, Nikolay Glushnev, Renshen Wang, Joshua Ainslie, Shangbang Long, Siyang Qin, Yasuhisa Fujii, Nan Hua, Tomas Pfister. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Scene text detection and document layout analysis have long been treated as two separate tasks in different image domains. In this paper, we bring them together and introduce the task of unified scene text detection and layout analysis. The first hierarchical scene text dataset is introduced to enable this novel research task. We also propose a novel method that is able to simultaneously detect scene text and form text clusters in a unified way. Comprehensive experiments show that our unified model achieves better performance than multiple well-designed baseline methods. Additionally, this model achieves state-of-the-art results on multiple scene text detection datasets without the need of complex post-processing. Dataset and code: https://github.com/google-research-datasets/hiertext.

We propose an end-to-end trainable network that can simultaneously detect and recognize text of arbitrary shape, making substantial progress on the open problem of reading scene text of irregular shape. We formulate arbitrary shape text detection as an instance segmentation problem; an attention model is then used to decode the textual content of each irregularly shaped text region without rectification. To extract useful irregularly shaped text instance features from image scale features, we propose a simple yet effective RoI masking step. Additionally, we show that predictions from an existing multi-step OCR engine can be leveraged as partially labeled training data, which leads to significant improvements in both the detection and recognition accuracy of our model. Our method surpasses the state-of-the-art for end-to-end recognition tasks on the ICDAR15 (straight) benchmark by 4.6%, and on the Total-Text (curved) benchmark by more than 16%.

We present a study with seven blind participants using three different mobile OCR apps to find text posted in various indoor environments. The first app considered was Microsoft SeeingAI in its Short Text mode, which reads any text in sight with a minimalistic interface. The second app was Spot+OCR, a custom application that separates the task of text detection from OCR proper. Upon detection of text in the image, Spot+OCR generates a short vibration; as soon as the user stabilizes the phone, a high-resolution snapshot is taken and OCR-processed. The third app, Guided OCR, was designed to guide the user in taking several pictures in a 360° span at the maximum resolution available by the camera, with minimum overlap between pictures. Quantitative results (in terms of true positive ratios and traversal speed) were recorded. Along with the qualitative observation and outcomes from an exit survey, these results allow us to identify and assess the different strategies used by our participants, as well as the challenges of operating these systems without sight.

We present a novel algorithm for geometry and camera pose reconstruction from image sequences that is specialized for indoor Manhattan scenes. Unlike general-purpose SfM/SLAM, our system represents geometric primitives in terms of canonically oriented planes. The algorithm starts by computing multi-planar segmentation and motion estimation from image pairs using constrained homographies. It then proceeds to recover the relative scale at each frame and to determine chains of match clusters, where each cluster is associated with a plane in the scene. Motion and scene geometry (expressed in terms of planar models) are then optimized using a novel formulation of Bundle Adjustment. Compared with other state-of-the-art SfM/SLAM algorithms, our technique is shown to produce superior and realistic surface reconstruction for a monocular indoor scene.

We propose a new technique for the accurate segmentation of text strokes from an image. The algorithm takes in a cropped image containing a word. It first performs a coarse segmentation using a Fully Convolutional Network (FCN). While not accurate, this initial segmentation can usually identify most of the text stroke content even in difficult situations, with uneven lighting and non-uniform background. The segmentation is then refined using a fully connected Conditional Random Field (CRF) with a novel kernel definition that includes stroke width information. In order to train the network, we created a new synthetic data set with 100K text images. Tested against standard benchmarks with pixellevel annotation (ICDAR 2003, ICDAR 2011, and SVT) our algorithm outperforms the state of the art by a noticeable margin.

We introduce an algorithm for word-level text spotting that is able to accurately and reliably determine the bounding regions of individual words of text "in the wild". Our system is formed by the cascade of two convolutional neural networks. The first network is fully convolutional and is in charge of detecting areas containing text. This results in a very reliable but possibly inaccurate segmentation of the input image. The second network (inspired by the popular YOLO architecture) analyzes each segment produced in the first stage, and predicts oriented rectangular regions containing individual words. No post-processing (e.g. text line grouping) is necessary. With execution time of 450 ms for a 1000 × 560 image on a Titan X GPU, our system achieves good performance on the ICDAR 2013, 2015 benchmarks [2], [1].

Selfies have become commonplace. More and more people take pictures of themselves, and enjoy enhancing these pictures using a variety of image processing techniques. One specific functionality of interest is automatic skin and hair segmentation, as this allows for processing one's skin and hair separately. Traditional approaches require user input in the form of fully specified trimaps, or at least of “scribbles” indicating foreground and background areas, with high-quality masks then generated via matting. Manual input, however, can be difficult or tedious, especially on a smartphone's small screen. In this paper, we propose the use of fully convolutional networks (FCN) and fully-connected CRF to perform pixel-level semantic segmentation into skin, hair and background. The trimap thus generated is given as input to a standard matting algorithm, resulting in accurate skin and hair alpha masks. Our method achieves state-of-the-art performance on the LFW Parts dataset [1]. The effectiveness of our method is also demonstrated with a specific application case.

We introduce an algorithm for text detection and localization ("spotting") that is computationally efficient and produces state-of-the-art results. Our system uses multi-channel MSERs to detect a large number of promising regions, then subsamples these regions using a clustering approach. Representatives of region clusters are binarized and then passed on to a deep network. A final line grouping stage forms word-level segments. On the ICDAR 2011 and 2015 benchmarks, our algorithm obtains an F-score of 82% and 83%, respectively, at a computational cost of 1.2 seconds per frame. We also introduce a version that is three times as fast, with only a slight reduction in performance.