Import my own dataset

[carmenrt]

Hello, I have tried to import my dataset through different ways but I have not been successful, could you tell me through which module or instruction I can import my own dataset? Thank you!

[BenjaminMidtvedt]

It depends on what you want to do, and the structure of your data.

Do you want to use the data to train a model, or do you want to get the predictions of a trained model on your data?
How is your data formatted? .tiff files? .png? All in the same folder, or different folders? Or on a remote server to download from?
Do you have annotations for your data? If so, what how are they formatted?

[carmenrt]

My idea is to use the data to train a cell detection and counting model. I currently have a folder containing training and test subfolders and subfolders containing images in .jpeg format and annotations in .json format.

[BenjaminMidtvedt]

Ok! How are the annotations encoded? As in, is it the centroid? A bounding box? Segmentations?

[carmenrt]

These are segmentations using COCO Annotator's polygon tool to select each of the cells I wanted to annotate within each image. Thanks!

[BenjaminMidtvedt]

Perfect!

We don't have anything as convenient as dt.LoadFromCOCO, but I found some old code on my computer for easily loading coco datasets. I'll paste it here; hopefully it can serve as a good baseline to create your pipeline.

import deeptrack as dt
import numpy as np
import imageio
import pycocotools.coco
import matplotlib.pyplot as plt

coco_json_file = r"path/to/coco/file.json"

# Load COCO annotations
coco = pycocotools.coco.COCO(coco_json_file)

# Get all image ids
image_ids = coco.getImgIds()

def coco_to_segmentation_mask(coco, image_id):
    annotation_ids = coco.getAnnIds(imgIds=image_id)
    annotations = coco.loadAnns(annotation_ids)
    masks = [coco.annToMask(annotation) for annotation in annotations]

    mask = np.any(masks, axis=0)
    return mask

def coco_to_image(coco, image_id):
    filename = coco.loadImgs(image_id)[0]["file_name"]
    root = os.path.dirname(coco_json_file)
    path = os.path.join(root, filename)
    image = imageio.imread(path)
    return image

root = dt.Arguments(
    image_id = lambda: int(np.random.choice(image_ids)),
)

image_loader = dt.Value(coco_to_image, coco=coco, image_id=root.image_id)
mask_loader = dt.Value(coco_to_segmentation_mask, coco=coco, image_id=root.image_id)

# augment image
image_pipeline = image_loader >> dt.NormalizeMinMax(0, 1) >> dt.Gaussian(sigma=0.05)

# Merge pipelines
image_and_mask_pipeline = image_pipeline & mask_loader

# Additional augmentations (Like flips, crops etc.)
...

# test pipeline
image_and_mask_pipeline.update()
image, mask = image_and_mask_pipeline()

# Plot image and mask
fig, ax = plt.subplots(1, 2, figsize=(10, 5))
ax[0].imshow(image)
ax[1].imshow(mask)
plt.show()

This is for training a very simple segmentation UNet. You may want to alter coco_to_segmentation_mask to suit your needs.