Skip to content

Latest commit

 

History

History
64 lines (47 loc) · 2.87 KB

assignment.md

File metadata and controls

64 lines (47 loc) · 2.87 KB

Assignment: Dataset Creation and Augmentation

1. Collect Two Suitable Videos

  • Capture two short videos using your camera or smartphone.
  • These videos will be used to create a training dataset for a binary classification model.
  • One video should contain positive examples, the other negative examples.
  • Move the camera while recording so that every frame is slightly different.
  • A few seconds of video are enough (aim for 100+ frames total).
  • Example:
    • Positive samples video: Camera circling around an apple (object) on a desk (background).
    • Negative samples video: Camera circling around an orange (object) on the same desk (background).

2. Convert Videos into Images

  • Extract every frame and save it as a .jpg image.
  • Resize the images to a reasonable size (e.g., 256×256).
  • Use a clear naming convention: image_<class>_<id>.jpg
  • <class>: positive / negative or 0 / 1
  • <id>: unique identifier for the frame

3. Create Training and Validation Split

  • Randomly split the images into two groups:
    • Training (~80%)
    • Validation (~20%)
  • Store them in separate folders for clarity.
  • Verify balance between positive and negative samples by visually checking a few images.

4. Suggest Augmentations for Your Data

  • Study possible augmentations in the Albumentations library.
  • Select reasonable augmentations that reflect real-world variations in your dataset.
  • Think carefully about transformations that make sense for your task (e.g., flipping, color changes, noise).

Expected Submission

Your final report should include:

  1. Dataset goal description

    • Define what counts as positive and negative samples.
    • Explain the purpose of your binary classification task.

  2. Video description

    • Provide details such as number of frames, resolution, objects, and background.

  3. Preview of images

    • Include thumbnails showing the diversity of both training and validation subsets.

  4. Chosen augmentations

    • List selected augmentations with before/after image examples.
    • Choose at least 5 augmentations.

  5. Conclusion

    • Explain why you chose certain augmentations and why you excluded others.

Hints

  • Example workflow notebook: dataset_creation.ipynb
  • Think carefully about the goal of your binary classification.
  • While selecting augmentations, ask yourself questions like:
    • Can the object realistically be flipped upside down or mirrored?
    • Can the object naturally vary in color?
    • Could there be lighting or brightness changes in real-world conditions?