Search

Highlights • We further develop the unravelled view of ResNets, which helps us better understand their behaviours. We demonstrate this in the context of a training process, which is the key difference from the original version 1. • We propose a group of relatively shallow convolutional networks based on our new understanding. Some of them perform comparably with the state-of-the-art approaches on the ImageNet classification dataset 2. • We evaluate the impact of using different networks on the performance of semantic image segmentation, and show these networks, as pre-trained features, can boost existing algorithms a lot. Abstract The community has been going deeper and deeper in designing one cutting edge network after another, yet some works are there suggesting that we may have gone too far in this dimension. Some researchers unravelled a residual network into an exponentially wider one, and assorted the success of residual networks to fusing a large amount of relatively shallow models. Since some of their early claims are still not settled, we in this paper dig more on this topic, i.e., the unravelled view of residual networks. Based on that, we try to find a good compromise between the depth and width. Afterwards, we walk through a typical pipeline of developing a deep-learning-based algorithm. We start from a group of relatively shallow networks, which perform as well or even better than the current (much deeper) state-of-the-art models on the ImageNet classification dataset. Then, we initialize fully convolutional networks (FCNs) using our pre-trained models, and tune them for semantic image segmentation. Results show that the proposed networks, as pre-trained features, can boost existing methods a lot. Even without exhausting the sophistical techniques to improve the classic FCN model, we achieve comparable results with the best performers on four widely-used datasets, i.e., Cityscapes, PASCAL VOC, ADE20k and PASCAL-Context. The code and pre-trained models are released for public access 1 1 https://github.com/itijyou/ademxapp. [ABSTRACT FROM AUTHOR]

Conditional Random Rields (CRF) have been widely applied in image segmentations. While most studies rely on hand-crafted features, we here propose to exploit a pre-trained large convolutional neural network (CNN) to generate deep features for CRF learning. The deep CNN is trained on the ImageNet dataset and transferred to image segmentations here for constructing potentials of superpixels. Then the CRF parameters are learnt using a structured support vector machine (SSVM). To fully exploit context information in inference, we construct spatially related co-occurrence pairwise potentials and incorporate them into the energy function. This prefers labelling of object pairs that frequently co-occur in a certain spatial layout and at the same time avoids implausible labellings during the inference. Extensive experiments on binary and multi-class segmentation benchmarks demonstrate the promise of the proposed method. We thus provide new baselines for the segmentation performance on the Weizmann horse, Graz-02, MSRC-21, Stanford Background and PASCAL VOC 2011 datasets. [ABSTRACT FROM AUTHOR]

• According to the detailed characteristics of the text instance, we dynamically generate the convolutional kernels from multi-feature for different instances. The specific attributes such as position, scale, and center, have been embedded into the convolutional kernel so that the mask prediction task using the text-instance-aware kernel will focus on the pixels that belong to themselves. Obviously, this design is helpful to improve the detection accuracy of adjacent text instances. • We generate the respective mask prediction head for each instance in parallel. These heads predict masks on the original feature map and retain resolution details of the text instance. It is no longer necessary to crop the RoIs and force them to be the same size. Our architecture overcomes the problem that a set of fixed convolution kernels cannot adapt to all resolutions, and at the same time preventing the loss of information caused by the multi-scales of the instances. • Because improving the text-instance-aware convolutional kernel increases the capacity of the model, we can also achieve competitive results with a very compact prediction head. Therefore, multiple mask prediction heads can be concurrently predicted without bringing significant computational overhead. • For the sake of improving the performance and accelerating the convergence of training, we design a text-shape sensitive position embedding to explicitly provide the location information to the mask prediction head. Arbitrarily shaped scene text detection has witnessed great development in recent years, and text detection using segmentation has been proven to an effective approach. However, problems caused by the diverse attributes of text instances, such as shapes, scales, and presentation styles (dense or sparse), persist. In this paper, we propose a novel text detector, termed DText, which can effectively formulate an arbitrarily shaped scene text detection task based on dynamic convolution. Our method can dynamically generate independent text-instance-aware convolutional parameters for each text instance from multi-features thus overcoming some intractable limitations of arbitrary text detection, such as the splitting of similar adjacent text, which poses challenges to fixed instance-shared convolutional parameters-based methods. Unlike standard segmentation methods relying on regions-of-interest bounding boxes, DText focuses on enhancing the flexibility of the network to retain details of instances from diverse resolutions while effectively improving prediction accuracy. Moreover, we propose encoding the shape and position information according to the characteristics of the text instance, termed text-shape sensitive position embedding. Thus, it can provide explicit shape and position information to the generator of the dynamic convolution parameters. Experiments on five benchmarks (Total-Text, SCUT-CTW1500, MSRA-TD500, ICDAR2015, and MLT) showed that our method achieves superior detection performance. [ABSTRACT FROM AUTHOR]