Introduction
Playing cards is a dataset created by Au-Zone for testing Object Detection algorithms at the edge. This dataset is continually being updated with more images and annotation to improve accuracy for prototyping. This dataset is useful for detecting playing cards in front of the camera. The values of each card are annotated (numbered cards plus face cards), however, suits are not currently annotated.
Use Case
Imagine a situation where a given user wants to detect small objects with rectangular shapes in front of the camera and the camera target distance is relatively small and all the objects have the same shape, but only the color and the texture is different among classes. Playing cards is the right dataset to test such a problem and the algorithm's performance since cards meet all the restrictions.
As a parallel use case, we can expect from a cards dataset that any algorithm that performs well will perform just as well in another problem with similar core restrictions. A real example of this could be to detect objects on a shelf. This last one has a high impact in retail businesses since automatic exploration, rotation, and counting objects in a shelf is very important to keep products up to date.
Dataset Stats
Playing cards has about 1476 annotated images, taken from different recording devices under different illumination conditions, orientations, and backgrounds. The dataset has 1328 training images and more than 4000 annotated boxes. The testing samples includes over 148 images and more than 140 annotation boxes.
Ace | Eight | Five | Four | Jack | King | Nine | Queen | Seven | Six | Ten | Three | Two | |
Train | 233 | 222 | 205 | 209 | 205 | 198 | 206 | 222 | 242 | 242 | 212 | 249 | 226 |
Test | 22 | 30 | 12 | 19 | 20 | 20 | 15 | 21 | 24 | 23 | 18 | 23 | 18 |
Table 1: Dataset distribution samples (Train images vs Test images)
Samples
The following are random samples from the dataset. All the cards are annotated
Evaluation Results
This dataset was intended to test Object Detection algorithms at the edge. For that reason, we summarize some results in Table 2 and Table 3.
Table 2: Evaluation metrics at (mAP @ 0.5). The following results were obtained after training ModelPack over 30 epochs.
Model | Float RTM | Per-Channel-Quantized RTM | Per-Tensor-Quantized RTM |
ModelPack | 93.21 % | 93.27 % | 92.48 % |
Table 3: NXP i.MX 8M Plus Inference timings on CPU, GPU and NPU (milliseconds)
Model | Float RTM | Per-Channel-Quantized RTM | Per-Tensor-Quantized RTM |
ModelPack-CPU | 491.61 | 264.84 | 264.71 |
ModelPack-GPU | 3231.72 | 2790.41 | 1063.74 |
ModelPack-NPU | 3830.98 | 17.99 |
14.47 |
Conclusions
In this article we have introduced our Playing Cards dataset as well as the results our ModelPack achieved after 57 training epochs.
Comments
0 comments
Please sign in to leave a comment.