{ "id": 319206, "title": "3rd Place --- Multi-scale detection", "url": "/competitions/ultra-mnist/discussion/319206", "author_name": "NamTran", "post_date": "2022-04-16T00:57:27.125000", "votes": 15, "comment_count": 0, "views": 0, "content": "

First and foremost, thank you especially to organizers who created an excited challenge.

\n

This is my first competition in kaggle playground, i'm very surprise when be 3rd winner :), So I'm honored to share my solution for this competition.

\n

Key-ideals

\n\n

Step By Step

\n
    \n
  1. I generated 60,000 images similar to the training and testing using the MNIST dataset.
    2. \n
  2. Training single fold yolov5x model with 1024 x 1012 solution (LB 0.88) - normal model
    3. \n
  3. Training another yolov5x tiny model to detect only 'tiny' digit in image (trained with special tiny dataset) - tiny model
    4. \n
  4. Repeat step 1, 2, 3 to get the best normal model and tiny model
    5. \n
  5. Deal with tiny digit in image by TTA (Test time augmentation) for boost accuracy of detectors.
    6. \n
  6. Ensemble result of predictions: 9 split, 16 split, 25 split, 768 x 768, 1024 x 1024, 1280 x 1280, 1536 x 1536 with list confidence_score [0.8, 0.8, 0.7, 0.95, 0.9, 0.9, 0.95]
    7. \n
\n

submission result: https://www.kaggle.com/code/namtran1118/ultramnist-3rd-solution

", "messages": [ { "id": 1756842, "postDate": "2022-04-16T00:57:27.127Z", "content": "

First and foremost, thank you especially to organizers who created an excited challenge.

\n

This is my first competition in kaggle playground, i'm very surprise when be 3rd winner :), So I'm honored to share my solution for this competition.

\n

Key-ideals

\n\n

Step By Step

\n
    \n
  1. I generated 60,000 images similar to the training and testing using the MNIST dataset.
    2. \n
  2. Training single fold yolov5x model with 1024 x 1012 solution (LB 0.88) - normal model
    3. \n
  3. Training another yolov5x tiny model to detect only 'tiny' digit in image (trained with special tiny dataset) - tiny model
    4. \n
  4. Repeat step 1, 2, 3 to get the best normal model and tiny model
    5. \n
  5. Deal with tiny digit in image by TTA (Test time augmentation) for boost accuracy of detectors.
    6. \n
  6. Ensemble result of predictions: 9 split, 16 split, 25 split, 768 x 768, 1024 x 1024, 1280 x 1280, 1536 x 1536 with list confidence_score [0.8, 0.8, 0.7, 0.95, 0.9, 0.9, 0.95]
    7. \n
\n

submission result: https://www.kaggle.com/code/namtran1118/ultramnist-3rd-solution

", "rawMarkdown": "First and foremost, thank you especially to organizers who created an excited challenge.\n\nThis is my first competition in kaggle playground, i'm very surprise when be 3rd winner :), So I'm honored to share my solution for this competition.\n\n**Key-ideals**\n+ object detection approach: detect all digit in image\n+ try to generate new dataset as closely as test dataset, and train object detector\n+ define batch_size_inference mechanism with yolov5 for TTA and speed-up inference time\n+ Source code for Generating dataset and training model is available here: https://github.com/NamTran11/UltraMNIST_3rd_solution,\n feel free to raise any issue for this code :)\n\n\n**Step By Step**\n1. I generated 60,000 images similar to the training and testing using the MNIST dataset.\n2. Training single fold yolov5x model with 1024 x 1012 solution (LB 0.88) - `normal model`\n3. Training another yolov5x `tiny` model to detect only 'tiny' digit in image (trained with special tiny dataset) - `tiny model`\n4. Repeat step 1, 2, 3 to get the best `normal model` and `tiny model`\n5. Deal with tiny digit in image by TTA (Test time augmentation) for boost accuracy of detectors.\n - with `normal model`: inference input image with multiple solutions, 768 x 768, 1024 x 1024, 1280 x 1280, 1536 x 1536\n - with `tiny model`: split input image in multiple sub-images, detect and merge result in each of sub-image with 9-split + 16-split + 25-split.\n6. Ensemble result of predictions: 9 split, 16 split, 25 split, 768 x 768, 1024 x 1024, 1280 x 1280, 1536 x 1536 with list confidence_score [0.8, 0.8, 0.7, 0.95, 0.9, 0.9, 0.95]\n\nsubmission result: https://www.kaggle.com/code/namtran1118/ultramnist-3rd-solution", "votes": 15 } ], "comments": [], "raw_markdown_by_id": { "1756842": "First and foremost, thank you especially to organizers who created an excited challenge.\n\nThis is my first competition in kaggle playground, i'm very surprise when be 3rd winner :), So I'm honored to share my solution for this competition.\n\n**Key-ideals**\n+ object detection approach: detect all digit in image\n+ try to generate new dataset as closely as test dataset, and train object detector\n+ define batch_size_inference mechanism with yolov5 for TTA and speed-up inference time\n+ Source code for Generating dataset and training model is available here: https://github.com/NamTran11/UltraMNIST_3rd_solution,\n feel free to raise any issue for this code :)\n\n\n**Step By Step**\n1. I generated 60,000 images similar to the training and testing using the MNIST dataset.\n2. Training single fold yolov5x model with 1024 x 1012 solution (LB 0.88) - `normal model`\n3. Training another yolov5x `tiny` model to detect only 'tiny' digit in image (trained with special tiny dataset) - `tiny model`\n4. Repeat step 1, 2, 3 to get the best `normal model` and `tiny model`\n5. Deal with tiny digit in image by TTA (Test time augmentation) for boost accuracy of detectors.\n - with `normal model`: inference input image with multiple solutions, 768 x 768, 1024 x 1024, 1280 x 1280, 1536 x 1536\n - with `tiny model`: split input image in multiple sub-images, detect and merge result in each of sub-image with 9-split + 16-split + 25-split.\n6. Ensemble result of predictions: 9 split, 16 split, 25 split, 768 x 768, 1024 x 1024, 1280 x 1280, 1536 x 1536 with list confidence_score [0.8, 0.8, 0.7, 0.95, 0.9, 0.9, 0.95]\n\nsubmission result: https://www.kaggle.com/code/namtran1118/ultramnist-3rd-solution" } }