Bug: LandmarkDataset raises image_loader attribute error when using store_imgs=False with in-memory images

[jejon]

Description:

When using the LandmarkDataset class from the landmarker package with store_imgs=False and providing in-memory images (not file paths) via imgs = [image], an unexpected AttributeError is raised:

AttributeError: 'LandmarkDataset' object has no attribute 'image_loader'

This is a bug — the dataset is expected to support store_imgs=False regardless of whether the input images are in-memory objects or file paths. The attribute image_loader is not defined anywhere in the code, making the source of this error unclear and unexpected.

---

Steps to Reproduce:

1. Set imgs = [image] where image is a loaded image (e.g., a PIL or NumPy object), not a file path

2. Initialize the dataset:

   ds = LandmarkDataset(imgs=imgs, store_imgs=False)

3. Use the dataset (e.g., iterate through it or index an item)

4. Observe the error:

   AttributeError: 'LandmarkDataset' object has no attribute 'image_loader'
   

---

Expected Behavior:

The dataset should support in-memory images with store_imgs=False and operate without errors.

---

Actual Behavior:

An AttributeError is thrown due to a missing image_loader attribute that is never explicitly defined or initialized.

---

Suggested Fix:

Ensure that LandmarkDataset handles in-memory images properly when store_imgs=False, and avoid referencing or relying on an undefined image_loader attribute. If image_loader is intended to be used internally, it should be properly initialized or the logic should be revised to avoid such dependency in this configuration.