Pre-Processing Results

Finetune and optimize your dataset with the help of the pre-processing results.

The AI Studio platform enables fashion users to create robust, production-ready AI models to address pressing fashion challenges. The process is iterative to ensure greater accuracy and performance of models.

As a result, all datasets go through a pre-processing where images are resized, poor quality images are removed, and data is prepared for building classifiers. Bounding boxes are added automatically to focus only on areas of interest. Through pose estimation, the data is prepared to focus only on relevant body parts while ignoring the background and noise.

The results of pre-processing enable the user to finetune and optimize the dataset by adding or removing fashion images. By using these insights, the user can ensure that the model is better trained to build classifiers and detectors, thereby, improving the model’s accuracy.

Images that cannot be used for training

When the images in the dataset are too small or corrupted, they are removed from the training dataset and categorized as ‘Images that cannot be used for training’.

(Insert Image)

Products that clashed with a better product

If an image in the training dataset has two or more fashion products, they are categorized as ‘Products that clashed with a better product’. For instance, the image could contain multiple people wearing adorning different fashion objects.

When products clash, the product occupying a larger area on the image is considered the better product and will, thus, be used in the dataset.

(Insert Image)

No fashion products found

Images that do not contain fashion objects of the selected categories are classified as ‘No Fashion Products found’.

For instance, the user selects topwear as the category while creating the project. But some of the images do not have any topwear or the topwear is not clearly visible. These images will be removed from the dataset.

(Insert Image)

Unnecessary Products/ Products are not needed for training

Images may contain multiple fashion objects. If one or more of those fashion products do not belong to the chosen categories, they are categorized as unnecessary products and will not be used for training.

For instance, the user has chosen bottomwear as the category while creating the project. However, in pre-processing other fashion objects such as dresses, topwear, footwear, etc. are detected in some of the images. These fashion objects are not required for training the AI model and thus, not be used.

(Insert Image)

Products that are manually removed

The user may choose to remove some of the products manually from the dataset. These are categorized as Products that are manually removed and will not be used for training.

(Insert Image)