For pipelines, benchmarking new models and various competitions usage of standardized computer vision datasets is a good idea. There are plenty of ways to design an image database according to your choice. You can also download or take screenshots of the picture you like the most on the web.
Once you have download images and label them in an excel spreadsheet, here are some tips to make the entire process busier and straightforward.
Steps needed once the download process gets completed
At first, you need the last number of images for search that is useful and relevant for your purpose. After this, you need to follow the steps, which are as follows
Filter out small images
When you download the images from the web, check the size of each image whether it Fits your purpose or not. The size of all images is not the same, so you need to filter out the images which are below a specific threshold. Generally, image models Take images ranges between 224×224 and 512×512; with the Help of the filter out option, you can cut the lower quality images.
Manual pruning
This feature allows removing known relevant or low-quality data from different phases of computer vision datasets. Once you finish the review mode, the images that you haven’t thrown out earlier will be left, so from here, you can just copy all these images into a new class that contains clear and quality images.
Remove duplicates
In your project, you will find plenty of similar duplicates; filter these duplicates with the help of resnet18. It is essential. Note that this feature does not make it practically possible on large computer vision datasets, but with 1-10K images, it is the best option to choose.
Labeling
By using the PyimageSearch method, you simply make multi-task problems, so it is necessary to mark different labels for each URLs set you have downloaded earlier. According to your project, you may also need to add some additional labels along with essential class names.
