Preparing & Uploading Data

Collect, format, and upload the images your model learns from

Overview

A good place to start when creating and training a new model is with a dataset containing at least 200 images — the console counts your uploads toward 200 and recommends starting there. Include diverse lighting, angles, and contexts for real-world applicability. The dataset should also include negative images. Negative images do not include the target object, and are helpful to reduce false positives in your trained model. Aim for roughly 80% of images showing your object and 20% negative images.

(Classification / Auto-Tag Models Only)

Tip: Prelabel Images by Folder Name

If your images are organized into folders named after their class labels (e.g., cat/, dog/, car/), you can drag and drop the folders directly into the upload interface to have the images automatically pre-tagged with the corresponding class.

For example:

training_images/
├── cat/
│   ├── image1.jpg
│   └── image2.jpg
├── dog/
│   └── image3.jpg

When you upload this structure:

• All images inside cat/ will be automatically tagged with the "cat" class

• All images inside dog/ will be tagged with "dog"

This saves time and ensures consistent labeling during dataset creation.

Note: This feature works best when folder names exactly match your intended class names.

What Are Negative Images?

Negative images are photos that do NOT contain your target object. These images are crucial for training your model to distinguish between what you want to detect and what you don't.

Why Use Negative Images?

• Reduce false positives: Help your model learn what NOT to classify as your target object

• Improve accuracy: Train the model to be more discriminating

• Handle similar objects: Teach the model to distinguish between similar-looking items

Examples of Negative Images:

• If training a door detection model: Include images of windows, archways, or openings that are NOT doors

• If training a car detection model: Include images of trucks, motorcycles, or street scenes without cars

• If training a person detection model: Include images of mannequins, statues, or scenes without people

How to Organize Your Dataset

Folder Structure

Organize your images into clear folder structures before uploading:

your-dataset/
├── positive-images/
│   ├── door_001.jpg
│   ├── door_002.jpg
│   └── ... (400 images of doors)
└── negatives/
    ├── window_001.jpg
    ├── archway_001.jpg
    └── ... (100 images without doors)

This layout is just for keeping track of your own files — folder names are not sent when you upload. What makes an image a negative is that you approve it during review with no boxes drawn on it. Images you leave unreviewed or reject are excluded from training entirely.

How to Upload Negative Images

Step-by-Step Process:

1. Prepare your negative images folder

   • Gather them somewhere separate so you can tell them apart from your positives

   • Add approximately 100 images that do NOT contain your target object

   • Ensure images show similar contexts or potentially confusing objects

2. Check that each image is one you want

   
3. Navigate to the upload interface

   • Go to My Models in the sidebar

   • Open your model

   • Continue to the upload step of the wizard

   
4. Upload the negative images

   • On the upload screen, locate the + Add more images tile

   

   • Drag and drop or upload your negative images onto this tile

   • Approve each one during review with no boxes drawn, and it trains as a negative

Upload Tips:

• You can drag and drop a whole folder at once

• The upload area will accept the folder and process all images inside

Best Practices for Negative Images

Quality Guidelines:

• Similar contexts: Use images from similar environments where your target object might appear

• Potential confusers: Include objects that might be mistaken for your target (like windows when training for doors)

• Varied conditions: Include different lighting, angles, and backgrounds

• High quality: Use clear, well-lit images similar to your positive examples

Common Mistakes to Avoid:

• Don't use completely unrelated images (random landscapes for a door model)

• Avoid low-quality or blurry negative images

• Don't include images that partially show your target object

Recommended Dataset Ratios

Dataset Size	Positive Images	Negative Images	Ratio
200 images	160 images	40 images	80/20
500 images	400 images	100 images	80/20
1000 images	800 images	200 images	80/20
2000+ images	1600+ images	400+ images	80/20

Frequently Asked Questions

Q: How do I upload negative images?

A: Gather images that don't contain your target object. Go to My Models → open your model → on the upload step, drag and drop them onto the + Add more images tile. Approve each one during review with no boxes drawn, and it trains as a negative.

Q: What should negative images contain?

A: Images that are similar to where your target object appears, but without the target object itself. For example, if training a door model, include images of windows, archways, or wall sections that might be confused for doors.

Q: How many negative images do I need?

A: Start with about 20% of your total dataset as negative images. For a 200-image dataset, use approximately 40 negative images.

Q: Can I add negative images after initial training?

A: Yes, you can upload additional negative images to improve your model's performance and reduce false positives.

---