Dataset Augmentation

Owned by Aniket Srinivasan Ashok
Last updated: 2024-06-27
2 min read

We augment the dataset to create more variation in the existing dataset (simulating images taken in varying conditions), and to increase the size of our training set itself.

Once all images have been labelled, navigate to Dataset in the left bar and ensure that the split is as close to 70/20/10 as possible (a few images of variance is fine). Then, navigate to Generate. Here, you will create a "Version" of the dataset.

Generate an augmented version:

  1. Leave default

  2. Leave default

  3. None

  4. Flip, Hue, Saturation, Brightness:

    1. (2022-2023 used the following:)

      1. Flip: Horizontal, Vertical 

      2. Hue: -27° and +27° 

      3. Saturation: -75% and +75% 

      4. Brightness: -25% and +0%

  5. 3x (but if a higher multiplier is available, use it)

  6. Click Generate (it takes a few minutes for the version to be generated)

To download a dataset version, see Dataset Download and Usage.

 

Note: Here are the settings used in 2023 for the competition:

image-20240627-145341.png

We do not know how well these augmentations worked for the team, since this wasn’t fully documented. Henceforth, we would like to experiment with various augmentations and document our results.

Suggestions for Data Augmentation

Roboflow documentation (Create Augmented Images | Roboflow Docs) suggests that changing the exposure, blur and noise be used with the augmentations used in 2023. They also have a detailed video on data augmentation for aerial data (Image Augmentations for Aerial Datasets).

If we have an excellent camera, noise may not be an issue, but a little blur could help replicate the quality of the drone’s camera. Blur would be useful to simulate conditions where objects are unclear from higher altitudes. As a reference point, if the dataset images are of identical quality to the landing pad images in Landing Pad Data Labelling Instructions, then blur and noise should be incorporated.

Roboflow also suggests cropping (Why and How to Implement Random Crop Data Augmentation) to help train the model in identifying various objects at different zooms (this includes cutting out part objects such as landing pads). We do not want partially cut images of these objects, so this augmentation is a no-go for now (if we have time, we could try to train a model capable of detecting sections of landing pads, but this may be unnecessary).

Use the 14-day free premium plan (no credit card needed) and we can use 5x instead of 3x as a dataset multiplier when creating our augmented dataset.

Augmentations to Experiment With

Note: update this list of augmentations when the ranges below are decided.

Augmentation 1 (replicating augmentations from 2023):

  • Flip: Horizontal, Vertical

  • Hue: -27° and +27° 

  • Saturation: -75% and 75%

  • Brightness: -25% and 0%

Augmentation 2:

  • Flip: Horizontal, Vertical

  • Hue: -27° to +27° [range]

  • Saturation: -75% to 75% [range]

  • Brightness: -25% to 0% [range]

  • Blur and Noise

Augmentation 3:

  • Flip: Horizontal, Vertical

  • Hue: -27° to +27° [range]

  • Saturation: -75% to 75% [range]

  • Brightness: -25% to 0% [range]

  • Crop, Blur and Noise

Past Data Augmentation Tasks

The Data Augmentation task from early 2024 can be found at 2024-04-03 Dataset Augmentation.

Â