{
"id": 148570,
"title": "Latest Detection Papers",
"url": "/competitions/global-wheat-detection/discussion/148570",
"author_name": "",
"post_date": "2020-05-04T20:35:03.647414500Z",
"votes": 22,
"comment_count": 1,
"views": 0,
"content": "\n\nModel efficiency has become increasingly important in\ncomputer vision. In this paper, we systematically study neural network architecture design choices for object detection and propose several key optimizations to improve efficiency. First, we propose a weighted bi-directional feature pyramid network (BiFPN), which allows easy and fast\nmulti-scale feature fusion; Second, we propose a compound\nscaling method that uniformly scales the resolution, depth,\nand width for all backbone, feature network, and box/class\nprediction networks at the same time. Based on these optimizations and EfficientNet backbones, we have developed\na new family of object detectors, called EfficientDet, which\nconsistently achieve much better efficiency than prior art\nacross a wide spectrum of resource constraints. In particular, with single-model and single-scale, our EfficientDetD7 achieves state-of-the-art 52.2 AP on COCO test-dev\nwith 52M parameters and 325B FLOPs1\n, being 4x – 9x\nsmaller and using 13x – 42x fewer FLOPs than previous detectors.\n![\"\"](\"https://media.arxiv-vanity.com/render-output/2547959/x1.png\")
\n\n\n\nThere are a huge number of features which are said to\nimprove Convolutional Neural Network (CNN) accuracy.\nPractical testing of combinations of such features on large\ndatasets, and theoretical justification of the result, is required. Some features operate on certain models exclusively\nand for certain problems exclusively, or only for small-scale\ndatasets; while some features, such as batch-normalization\nand residual-connections, are applicable to the majority of\nmodels, tasks, and datasets. We assume that such universal\nfeatures include Weighted-Residual-Connections (WRC),\nCross-Stage-Partial-connections (CSP), Cross mini-Batch\nNormalization (CmBN), Self-adversarial-training (SAT)\nand Mish-activation. We use new features: WRC, CSP,\nCmBN, SAT, Mish activation, Mosaic data augmentation,\nCmBN, DropBlock regularization, and CIoU loss, and combine some of them to achieve state-of-the-art results: 43.5%\nAP (65.7% AP50) for the MS COCO dataset at a realtime speed of ∼65 FPS on Tesla V100\n
\n\n\n I will keep updating this post with more relevant papers
\n
",
"messages": [
{
"id": "833465",
"postDate": "05/04/2020 20:35:03",
"content": "\n\nModel efficiency has become increasingly important in\ncomputer vision. In this paper, we systematically study neural network architecture design choices for object detection and propose several key optimizations to improve efficiency. First, we propose a weighted bi-directional feature pyramid network (BiFPN), which allows easy and fast\nmulti-scale feature fusion; Second, we propose a compound\nscaling method that uniformly scales the resolution, depth,\nand width for all backbone, feature network, and box/class\nprediction networks at the same time. Based on these optimizations and EfficientNet backbones, we have developed\na new family of object detectors, called EfficientDet, which\nconsistently achieve much better efficiency than prior art\nacross a wide spectrum of resource constraints. In particular, with single-model and single-scale, our EfficientDetD7 achieves state-of-the-art 52.2 AP on COCO test-dev\nwith 52M parameters and 325B FLOPs1\n, being 4x – 9x\nsmaller and using 13x – 42x fewer FLOPs than previous detectors.\n
\n\n\n\nThere are a huge number of features which are said to\nimprove Convolutional Neural Network (CNN) accuracy.\nPractical testing of combinations of such features on large\ndatasets, and theoretical justification of the result, is required. Some features operate on certain models exclusively\nand for certain problems exclusively, or only for small-scale\ndatasets; while some features, such as batch-normalization\nand residual-connections, are applicable to the majority of\nmodels, tasks, and datasets. We assume that such universal\nfeatures include Weighted-Residual-Connections (WRC),\nCross-Stage-Partial-connections (CSP), Cross mini-Batch\nNormalization (CmBN), Self-adversarial-training (SAT)\nand Mish-activation. We use new features: WRC, CSP,\nCmBN, SAT, Mish activation, Mosaic data augmentation,\nCmBN, DropBlock regularization, and CIoU loss, and combine some of them to achieve state-of-the-art results: 43.5%\nAP (65.7% AP50) for the MS COCO dataset at a realtime speed of ∼65 FPS on Tesla V100\n
\n\n\n I will keep updating this post with more relevant papers
\n
",
"rawMarkdown": "##1. [EfficientDet: Scalable and Efficient Object Detection](https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch)\n`Model efficiency has become increasingly important in\ncomputer vision. In this paper, we systematically study neural network architecture design choices for object detection and propose several key optimizations to improve efficiency. First, we propose a weighted bi-directional feature pyramid network (BiFPN), which allows easy and fast\nmulti-scale feature fusion; Second, we propose a compound\nscaling method that uniformly scales the resolution, depth,\nand width for all backbone, feature network, and box/class\nprediction networks at the same time. Based on these optimizations and EfficientNet backbones, we have developed\na new family of object detectors, called EfficientDet, which\nconsistently achieve much better efficiency than prior art\nacross a wide spectrum of resource constraints. In particular, with single-model and single-scale, our EfficientDetD7 achieves state-of-the-art 52.2 AP on COCO test-dev\nwith 52M parameters and 325B FLOPs1\n, being 4x – 9x\nsmaller and using 13x – 42x fewer FLOPs than previous detectors.`\n\n\n##2. [YOLOv4: Optimal Speed and Accuracy of Object Detection](https://github.com/AlexeyAB/darknet)\n`There are a huge number of features which are said to\nimprove Convolutional Neural Network (CNN) accuracy.\nPractical testing of combinations of such features on large\ndatasets, and theoretical justification of the result, is required. Some features operate on certain models exclusively\nand for certain problems exclusively, or only for small-scale\ndatasets; while some features, such as batch-normalization\nand residual-connections, are applicable to the majority of\nmodels, tasks, and datasets. We assume that such universal\nfeatures include Weighted-Residual-Connections (WRC),\nCross-Stage-Partial-connections (CSP), Cross mini-Batch\nNormalization (CmBN), Self-adversarial-training (SAT)\nand Mish-activation. We use new features: WRC, CSP,\nCmBN, SAT, Mish activation, Mosaic data augmentation,\nCmBN, DropBlock regularization, and CIoU loss, and combine some of them to achieve state-of-the-art results: 43.5%\nAP (65.7% AP50) for the MS COCO dataset at a realtime speed of ∼65 FPS on Tesla V100`\n\n\n>I will keep updating this post with more relevant papers",
"votes": null
},
{
"id": "834180",
"postDate": "05/05/2020 11:16:54",
"content": "Paper on Detection and analysis of wheat spikes using Convolutional Neural Networks
",
"rawMarkdown": "Paper on [Detection and analysis of wheat spikes using Convolutional Neural Networks](https://plantmethods.biomedcentral.com/articles/10.1186/s13007-018-0366-8)",
"votes": null
}
],
"comments": [
{
"id": 834180,
"author_name": "prateek0x",
"author_url": "",
"post_date": "05/05/2020 11:16:54",
"content": "Paper on Detection and analysis of wheat spikes using Convolutional Neural Networks
",
"votes": null,
"replies": []
}
],
"raw_markdown_by_id": {
"833465": "##1. [EfficientDet: Scalable and Efficient Object Detection](https://github.com/zylo117/Yet-Another-EfficientDet-Pytorch)\n`Model efficiency has become increasingly important in\ncomputer vision. In this paper, we systematically study neural network architecture design choices for object detection and propose several key optimizations to improve efficiency. First, we propose a weighted bi-directional feature pyramid network (BiFPN), which allows easy and fast\nmulti-scale feature fusion; Second, we propose a compound\nscaling method that uniformly scales the resolution, depth,\nand width for all backbone, feature network, and box/class\nprediction networks at the same time. Based on these optimizations and EfficientNet backbones, we have developed\na new family of object detectors, called EfficientDet, which\nconsistently achieve much better efficiency than prior art\nacross a wide spectrum of resource constraints. In particular, with single-model and single-scale, our EfficientDetD7 achieves state-of-the-art 52.2 AP on COCO test-dev\nwith 52M parameters and 325B FLOPs1\n, being 4x – 9x\nsmaller and using 13x – 42x fewer FLOPs than previous detectors.`\n\n\n##2. [YOLOv4: Optimal Speed and Accuracy of Object Detection](https://github.com/AlexeyAB/darknet)\n`There are a huge number of features which are said to\nimprove Convolutional Neural Network (CNN) accuracy.\nPractical testing of combinations of such features on large\ndatasets, and theoretical justification of the result, is required. Some features operate on certain models exclusively\nand for certain problems exclusively, or only for small-scale\ndatasets; while some features, such as batch-normalization\nand residual-connections, are applicable to the majority of\nmodels, tasks, and datasets. We assume that such universal\nfeatures include Weighted-Residual-Connections (WRC),\nCross-Stage-Partial-connections (CSP), Cross mini-Batch\nNormalization (CmBN), Self-adversarial-training (SAT)\nand Mish-activation. We use new features: WRC, CSP,\nCmBN, SAT, Mish activation, Mosaic data augmentation,\nCmBN, DropBlock regularization, and CIoU loss, and combine some of them to achieve state-of-the-art results: 43.5%\nAP (65.7% AP50) for the MS COCO dataset at a realtime speed of ∼65 FPS on Tesla V100`\n\n\n>I will keep updating this post with more relevant papers",
"834180": "Paper on [Detection and analysis of wheat spikes using Convolutional Neural Networks](https://plantmethods.biomedcentral.com/articles/10.1186/s13007-018-0366-8)"
},
"source": "meta"
}