Skip to main content
Back

PINTO0309 Converts DDN Edge Detector to Run on 8 Deployment Frameworks

PINTO0309 converts DDN, a state-of-the-art edge detector, to ONNX and multi-framework formats for immediate edge deployment

Repo, Full Access
  • DDN converted to ONNX: PINTO0309 exported the Doubly Decoupled Network edge detector to ONNX and 8+ deployment formats.
  • State-of-the-art accuracy: DDN scores ODS 0.867 on BSDS500 at VGG-level compute cost, accepted in Neurocomputing.
  • Two-trick architecture: DDN uses data decoupling (learnable Gaussians) and feature decoupling (spatial/semantic split) to cut overfitting and redundant compute.
  • Instant deployment: Pre-built weights on S3, download scripts included — works on Coral EdgeTPU, OpenVINO, CoreML, TFLite, and browsers.
  • Broader use: DDN also transfers to retinal vessel segmentation and salient object detection beyond classical edge maps.
  • PINTO Model Zoo now holds 480+ models across 8 frameworks with quantization variants per model.

Katsuya Hyodo (PINTO0309), the one-person operation behind the PINTO Model Zoo, has converted DDN , a Doubly Decoupled Network for edge detection , from its original PyTorch form into ONNX and a full suite of deployment-ready formats. The conversion lands DDN into a zoo that now holds 480+ models, each pre-packaged for immediate use across hardware from server GPUs to microcontrollers.

What DDN actually does

Edge detection , finding the boundaries between objects in an image , sounds simple but is notoriously hard to scale. Deep learning-based edge detectors are increasingly scaled up to achieve higher performance, but this growth comes at the cost of higher computational demands and a greater risk of model overfitting.

DDN alleviates the model overfitting problem by data decoupling, and compresses the redundant computation in shallow features by feature decoupling. The "doubly" in the name refers to two distinct decoupling moves the architecture makes:

  • Data decoupling: Image-edge pairs are decoupled and mapped to learnable Gaussian distributions, with corresponding edges mapped to samples of those distributions , generating more robust features by normalizing latent variables and sampling from the Gaussian distributions.
  • Feature decoupling: Image features are split into spatial and semantic streams and encoded separately, so shallow features no longer need to provide cues for semantic features and their diversity can be drastically compressed.