{
  "id": 302231,
  "title": "Testing the limits of YOLO V1's performance",
  "url": "/competitions/tensorflow-great-barrier-reef/discussion/302231",
  "author_name": "Adnan Pen",
  "post_date": "2022-01-21T13:46:38.147000",
  "votes": 2,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Recently implemented <a href=\"https://www.kaggle.com/adnanpen/yolov1-from-scratch-for-cots\" target=\"_blank\">YOLO V1  from scratch</a> and started training it on the COTS dataset without using any prior weights to fine-tuning on. However, going through the paper while implementing the model I came to know that the authors first fine-tuned the <strong>Darknet model</strong> on ImageNet for classification and then increased the input size when fine-tuning it on the COCO dataset for object detection to further increase the accuracy.</p>\n<p>What I'd like to know is how we can <strong>improve the accuracy of the vanilla YOLO v1 architecture without making changes to the original model</strong> <br>\nRecently there have been several papers that take up traditional architectures and apply modern techniques to them, which has resulted in equivalent output in performance from traditional architectures when compared to modern architectures. Examples include ResNet strikes back, Convnext, MLP-mixer etc.</p>\n<p>In a similar manner are there any techniques/tweaks/tricks that can Improve accuracy for Object Detection without making changes to the architecture?<br>\n<strong><em>Think of it like creating YOLO V1.1</em></strong>🔥</p>",
  "messages": [
    {
      "id": 1659016,
      "postDate": "2022-01-21T13:46:38.147Z",
      "content": "<p>Recently implemented <a href=\"https://www.kaggle.com/adnanpen/yolov1-from-scratch-for-cots\" target=\"_blank\">YOLO V1  from scratch</a> and started training it on the COTS dataset without using any prior weights to fine-tuning on. However, going through the paper while implementing the model I came to know that the authors first fine-tuned the <strong>Darknet model</strong> on ImageNet for classification and then increased the input size when fine-tuning it on the COCO dataset for object detection to further increase the accuracy.</p>\n<p>What I'd like to know is how we can <strong>improve the accuracy of the vanilla YOLO v1 architecture without making changes to the original model</strong> <br>\nRecently there have been several papers that take up traditional architectures and apply modern techniques to them, which has resulted in equivalent output in performance from traditional architectures when compared to modern architectures. Examples include ResNet strikes back, Convnext, MLP-mixer etc.</p>\n<p>In a similar manner are there any techniques/tweaks/tricks that can Improve accuracy for Object Detection without making changes to the architecture?<br>\n<strong><em>Think of it like creating YOLO V1.1</em></strong>🔥</p>",
      "rawMarkdown": "Recently implemented [YOLO V1  from scratch](https://www.kaggle.com/adnanpen/yolov1-from-scratch-for-cots) and started training it on the COTS dataset without using any prior weights to fine-tuning on. However, going through the paper while implementing the model I came to know that the authors first fine-tuned the **Darknet model** on ImageNet for classification and then increased the input size when fine-tuning it on the COCO dataset for object detection to further increase the accuracy.\n\nWhat I'd like to know is how we can **improve the accuracy of the vanilla YOLO v1 architecture without making changes to the original model** \nRecently there have been several papers that take up traditional architectures and apply modern techniques to them, which has resulted in equivalent output in performance from traditional architectures when compared to modern architectures. Examples include ResNet strikes back, Convnext, MLP-mixer etc.\n\nIn a similar manner are there any techniques/tweaks/tricks that can Improve accuracy for Object Detection without making changes to the architecture?\n***Think of it like creating YOLO V1.1***🔥",
      "votes": 2
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1659016": "Recently implemented [YOLO V1  from scratch](https://www.kaggle.com/adnanpen/yolov1-from-scratch-for-cots) and started training it on the COTS dataset without using any prior weights to fine-tuning on. However, going through the paper while implementing the model I came to know that the authors first fine-tuned the **Darknet model** on ImageNet for classification and then increased the input size when fine-tuning it on the COCO dataset for object detection to further increase the accuracy.\n\nWhat I'd like to know is how we can **improve the accuracy of the vanilla YOLO v1 architecture without making changes to the original model** \nRecently there have been several papers that take up traditional architectures and apply modern techniques to them, which has resulted in equivalent output in performance from traditional architectures when compared to modern architectures. Examples include ResNet strikes back, Convnext, MLP-mixer etc.\n\nIn a similar manner are there any techniques/tweaks/tricks that can Improve accuracy for Object Detection without making changes to the architecture?\n***Think of it like creating YOLO V1.1***🔥"
  }
}