Reproduce “DSSD : Deconvolutional Single Shot Detector” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Cheng-Yang Fu, Wei Liu, Ananth Ranga, Ambrish Tyagi, Alexander C. Berg

The Deconvolutional Single Shot Detector (DSSD) from UNC Chapel Hill and Amazon Inc. extends the SSD framework by integrating a Residual-101 backbone and an asymmetric deconvolutional architecture for enhanced contextual information. The DSSD 513 model achieved 81.5% mAP on PASCAL VOC2007 and 33.2% mAP on COCO test-dev2015, showing improved accuracy, particularly for small objects, compared to earlier single-shot methods.

Key insights

  • Simply replacing a backbone with a deeper network like Residual-101 is insufficient for single-shot detectors; specific architectural adaptations, such as a prediction module, are crucial for performance.
  • An asymmetric encoder-decoder structure, combining a deep encoder with a shallower deconvolutional decoder and element-wise product skip connections, effectively integrates high-level semantic context with fine-grained spatial details, improving small object detection.
  • The method of fusing features (element-wise product vs. sum) in deconvolutional skip connections significantly impacts performance, with product fusion proving more effective for combining different feature modalities.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.