SKU110K-DenseDet : A Machine Learning Model That Can Detect Products in a Store.
This is an introduction to「SKU110K-DenseDet」, a machine learning model that can be used with ailia SDK. You can easily use this model to create AI applications using ailia SDK as well as many other ready-to-use ailia MODELS.
SKU110K-DenseDet is a machine learning model that can detect products in a store. It can detect the bounding boxes of products presented on a supermarket shelf, but there is no categorization, only the presence of a product is detected.
A Solution to Product detection in Densely Packed Scenes
This work is a solution to densely packed scenes dataset SKU-110k. Our work is modified from Cascade R-CNN. To solve…
GitHub — Media-Smart/SKU110K-DenseDet: A state of art detector for densely packed scenes dataset…
A state of art detector for densely packed scenes dataset SKU-110K. For more information, please read our technical…
SKU110K is a data set for product detection published in April 2019 which contains 11,762 images taken with cell phones in thousands of supermarkets around the world (United States, Europe, East Asia). Bounding boxes were manually annotated. It contains 90,968 bounding boxes in 8,233 images for training and 432,312 bounding boxes in 2,941 images for validation.
GitHub — eg4000/SKU110K_CVPR19
Dataset and Codebase for CVPR2019 “Precise Detection in Densely Packed Scenes” [Paper link] A typical image in our…
Precise Detection in Densely Packed Scenes
Man-made scenes can be densely packed, containing numerous objects, often identical, positioned in close proximity. We…
SKU110K-DenseDet was trained using MMDetection. It achieves a 58.0% mAP using Cascade R-CNN. The backbone uses ResNXt-101.
Since SKU110K contains many small objects, usual architectures and input resolutions for object detection is not accurate enough. Therefore input size of images is set to 2560x2560.
In environments with low GPU memory, random cropping of images is used during training, in a way that it does not negatively impact the training results.
SKU110K contains on average 150 bounding boxes per image. This is significantly more than MS COCO and default hyper parameters are not optimal. Hence the max positive sample number of both RPN and R-CNN sampler were adjusted to release the limits.
SKU110K-DenseDet can be executed using ailia SDK with the following command. Due to the huge size of the backbone, please specify
-e 0 to run in CPU mode in environments with low GPU memory.
$ python3 sku110k-densedet.py -i input.jpg -e 0