Pooling Pyramid Network for Object Detection
This work addresses model efficiency for object detection applications, but it is incremental as it builds on existing SSD methods.
The paper tackles the problem of reducing model size in object detection by proposing a tweak to the Single Shot Multibox Detector (SSD) family, which shares box predictors across scales and replaces convolutions with max pooling, resulting in a significantly smaller model while maintaining the same quality as vanilla SSD.
We'd like to share a simple tweak of Single Shot Multibox Detector (SSD) family of detectors, which is effective in reducing model size while maintaining the same quality. We share box predictors across all scales, and replace convolution between scales with max pooling. This has two advantages over vanilla SSD: (1) it avoids score miscalibration across scales; (2) the shared predictor sees the training data over all scales. Since we reduce the number of predictors to one, and trim all convolutions between them, model size is significantly smaller. We empirically show that these changes do not hurt model quality compared to vanilla SSD.