{
  "fetched_at": "2026-10-19T06:31:33Z",
  "source_url": "https://www.kaggle.com/competitions/bugnist2024fgvc/data",
  "pages": [
    {
      "name": "data-description",
      "content": "# BugNIST Dataset and Submission Guidelines\n\nThe [BugNIST](https://arxiv.org/abs/2304.01838) dataset is specifically designed for the task of volumetric object detection under domain shift. The training data consists of volumes of individual bugs, while the test data is a mixture of bugs and other materials such as leaves, wood shavings, cotton, and shredded paper.\n\n## Dataset Structure\n\n### Training Data\nThe training dataset is located in the `BugNIST_DATA/train` folder. It is divided into 12 subfolders, each containing single-volume scans of a specific class of bug, as indicated by the folder name. Each volume measures 128-by-64-by-64 voxels (z-axis corresponds to the longest axis). For submission the abbreviated name should be used.\n\nHere is a breakdown of the individual volumes:\n\n| Abbreviation     | **AC**        | **BC**        | **BF**   | **BL**        | **BP**            | **CF**         |\n|------------------|---------------|---------------|----------|---------------|-------------------|----------------|\n| Bug Type         | Brown Cricket | Black Cricket | Blowfly  | Beetle Larvae | Blowfly Pupae     | Curly-wing Fly |\n| # of Volumes     | 724           | 762           | 734      | 780           | 752               | 756            |\n| **Abbreviation** | **GH**        | **MA**        | **ML**   | **PP**        | **SL**            | **WO**         |\n| Bug Type         | Grasshopper   | Maggot        | Mealworm | Fly Pupae     | Soldierfly Larvae | Woodlice       |\n| # of Volumes     | 716           | 765           | 738      | 769           | 740               | 949            |\n\n### Validation Data\nThe validation dataset is located in the `BugNIST_DATA/validation` folder. It contains 78 volumes of bug mixtures in `.tif` file format. The dimensions are (128, 92, 92) corresponding to the z,x,y-axis. The true labels are provided in the `validation.csv` file within the same folder.\n\n### Test Data\nThe test dataset is located in the `BugNIST_DATA/test` folder. It contains 310 volumes of bug mixtures in `.tif` file format. The dimensions are (128, 92, 92) corresponding to the z,x,y-axis. For submission of predictions on the test set, use the `Sample_solution.csv` file.\n\n## Submission Format\nTo submit predictions for evaluation on the validation data and for participation in the Kaggle competition, you must adhere to a specific format. Use the provided `Sample_solution.csv` file as a reference for formatting your predictions.\n\nYour prediction file should be a CSV with two columns: 'filename' and 'centerpoints'.\n\n- 'filename': This column should contain the name of the volume file. This is already given in the file `Sample_solution.csv`.\n- 'centerpoints': This column should contain the centerpoint and class prediction for each bug detected in the volume.\n\nThe format for each bug's centerpoint is as follows:\n- Class label (as given in the table above), followed by a semicolon (;)\n- x-coordinate of the centerpoint, followed by a semicolon (;)\n- y-coordinate of the centerpoint, followed by a semicolon (;)\n- z-coordinate of the centerpoint\n\nIf multiple bugs are detected in the same volume, their centerpoints should be combined in the same cell, separated by semicolons (;).\n\nFor example, if a bug of class \"Beetle Larva\" (\"bl\") is detected with a centerpoint at x-coordinate: 48.61, y-coordinate: 39.14, and z-coordinate 96.19, it should be written as a string \"bl;48.61;39.14;96.19\". If an additional bug of type \"Soldier Fly Larva\" (\"sl\") is detected at position (77.71, 53.14, 57.17), it can be combined in the cell as a string \"bl;48.61;39.14;96.19;sl;77.71;53.14;57.17\".\n\nHere is an example of a submission file:\n\n| Filename                    | Centerpoints                                      |\n| --------------------------- | ------------------------------------------------- |\n| mix_02_006_centroids.csv    | bl;48.61;39.14;96.19;sl;77.71;53.14;57.17        |\n| mix_02_011_centroids.csv    | ml;77.86;46.26;57.79;sl;41.14;61.93;32.37;sl;56.09;48.52;56.20 |\n\nEnsure your submission file follows this format for proper evaluation and scoring using the F1 class label score.\n\nThe notebook `BugNIST_example_load.ipynb` contains an example load function. It loads an image and a string of formatted centerpoints which it uses to display the centerpoints in a figure.\n\nNb. z-coordinate corresponds to the longest axis i.e. we expect a center coordinate value between 0 and 128. Both the x and y-axis have dimensions of 92.\n\n## Evaluation \nPredictions are evaluated by matching the predicted centerpoints to the true centerpoints. This is done by calculating the `scipy.spatial.distance.cdist` and then matching the points using `scipy.optimize.linear_sum_assignment`. The labels of the matched points are then compared to calculate the F1 score.\n\nThe evaluation metric is provided in the `BugNIST_metric.py` file. You can use this to locally evaluate the validation data.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Description",
      "content": "![](https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F18910303%2F86b7c7c4511352005f3b429a657ead89%2Fillustration2.png?generation=1709301396521694&alt=media)\nWe provide the scans of individual bugs for training, a validation set of mixtures with annotated center point coordinates and bug type, and two test sets of mixtures without annotations for the public and private leaderboards. Your task is to build a model that predicts the center points and bug type for the mixed volumes.\n\nFor training the model, the total number of volumes of individual bugs is 9185. There are 388 mixtures, where a validation set of 78 is provided with center point annotation and 155 mixture volumes are for the public and private leader boards respectively. \n\n\n\n",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "Evaluation",
      "content": "See \"Data\" section for overview over submission format and evaluation metric.\n",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "rules",
      "content": "Please follow all rules for the FGVC11 Workshop.\n\nIn addition, you should train your model only on the provided training set containing scans of individual bugs. The validation set with the scans of bug mixtures is provided only for validating your predictions and making sure that you comply with the submission template. Therefore, you are **not** allowed to train on the validation data of mixtures. The BugNIST2024 competition evaluates domain shift, and this will not be the case if you use the validation data while training.\n\nYou are free to make synthetic mixtures using the provided training set of individually scanned bugs. This can be a useful data augmentation for training your model.",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    },
    {
      "name": "abstract",
      "content": "##Goal:\nWe welcome you to the BugNIST challenge, which is part of the FGVC11 workshop at CVPR 2024. Here you will compete in detecting and classifying objects in volumes under domain shift and hereby contribute to advancing methods for the analysis of volumetric data. Based on volumes containing one bug each, your task is to train a model that can detect and classify bugs in volumes that contain several bugs mixed with other materials. \n\n##Collaborative paper!\nAfter the FGVC workshop, we will collect short descriptions from the participating teams to write a collaborative paper on approaches for detection under domain shift. We will reach out to you for your input.\n\n##Baseline implementation!\nThere is now a baseline implementation for detection available. See the *U-Net baseline* under the *Code* tab.\n\n##Why does it matter?\nIn volumetric imaging, it is often necessary to detect and classify objects that are densely packed in a mixture of other objects or materials that are not of interest. Annotating such data is tedious and difficult, and instead, it is much easier to scan some representative examples of the type of objects that should be detected. The problem is that individually scanned objects are placed in a context that is very different from the mixed objects that we wish to detect. This is a domain shift, where the objects of interest have the same appearance, but the context surrounding the objects is different.\n\nTo promote the development of new methods for domain adaptation in volumetric scans, we have created the BugNIST dataset, which is the basis for the BugNIST challenge. The dataset is based on 12 types of bugs such as larvae, pupae, grasshoppers, etc. The training data is more than 700 micro-CT scans of individual bugs of each type, whereas the test data is mixtures. \n\n",
      "isPublished": false,
      "hasIsPublished": false,
      "postTitle": "",
      "hasPostTitle": false,
      "mimeType": "",
      "hasMimeType": false
    }
  ]
}