The Canned Goods Dataset

by Vizonix

This dataset differentiates between 4 similar object classes: 4 types of canned goods. We built this dataset with cans of olives, beans, stewed tomatoes, and refried beans.

The dataset is pre-augmented. That is to say, all required augmentations are applied to the actual native dataset prior to inference. We have found that augmenting this way provides our users maximum visibility and flexibility in tuning their dataset (and classifier) to achieve their specific use-case goals. Augmentations are present and visible in the native dataset prior to the classifier - so it's never a mystery what augmentation tweaks produce a more positive or negative outcome during training. It also eliminates the risk of downsizing affecting annotations.

The training images in this dataset were created in our studio in Florida from actual physical objects to the following specifications:

• Each item was imaged using a 360-degree horizontal rotation - imaged every 9 degrees at 0 degrees of elevation.
• Each item was imaged (per above) 3 times - using physical left lighting, right lighting, and frontal lighting.
• Backgrounds in this dataset are completely random - they do not factor into the classifier's decision-making (nor do we ever want them to). We used 100% random backgrounds generated in-house. This eliminates background bias in the dataset. Our use of random backgrounds is a newly released feature in our datasets.

The training images in this dataset were composited / augmented in this way:

• Imaged objects were randomly rotated in frame from 15 to 340 degrees.
• Imaged objects were randomly positioned in frame.
• Imaged objects were randomly sized from .33 to 1.0 of original.
• Image contrast was randomly adjusted from .7 to 1.25.
• Gaussian blur was randomly introduced at a factor from 2 to 5.
• Color channels were dropped randomly (R,G,B).
• Grayscale images were introduced randomly.
• Soft occlusions (noise, and others) in random transperencies were randomly introduced.
• Hard occlusions (noise and others) in solid transperencies were randomly introduced.
• Brightness was randomly adjusted.
• Sharpness was randomly adjusted.
• Color balance was randomly adjusted.
• Images were resized to 640x640 for Roboflow's platform.

1,600 (+) different images were uploaded for each class (out of the 25,000 total images created for each class).

Understanding our Dataset Insights File

As users train their classifiers, they often wish to enhance accuracy by experimenting with or tweaking their dataset. With our Dataset Insights documents, they can easily determine which images possess which augmentations. Dataset Insights allow users to easily add or remove images with specific augmentations as they wish. This also provides a detailed profile and inventory of each file in the dataset.

The Dataset Insights document enables the user to see exactly which source image, angle, augmentation(s), etc. were used to create each image in the dataset.

Dataset Insight Files:

• Beans: [https://176288177.s3.amazonaws.com/Dataset_Insights_Beans.xlsx]
• Olives: [https://176288177.s3.amazonaws.com/Dataset_Insights_Olives.xlsx]
• Refried Beans: [https://176288177.s3.amazonaws.com/Dataset_Insights_Refried_Beans.xlsx]
• Tomatoes: [https://176288177.s3.amazonaws.com/Dataset_Insights_Tomatoes.xlsx]

About Vizonix

Vizonix (vizonix.com) creates from-scratch datasets created from 100% in-house generated photography. Our images and backgrounds are generated in-house in our Florida studio. We typically image smaller items, deliver in 72 hours, and specialize in Manufacturer Quality Assurance (MQA) datasets.