What it does CTPN (Connectionist Text Proposal Network) finds horizontal text lines in scene images—signs, posters, whatever text appears in the wild. It outputs bounding boxes without the “side-refinement” step mentioned in the original ECCV paper. The authors also run an online demo at textdet.com.

The interesting bit The repo bundles its own Caffe fork with three custom layers (Reverse, Transpose, LSTM) that you’d need to port if you wanted to use a different Caffe build. That’s unusual even for 2016—most projects just listed layer definitions. The authors also ship a 78MB pretrained model, so you can run the demo without training yourself.

Key highlights

• Detects text lines in natural scene images (not just scanned documents)
• Ships with pretrained model and custom Caffe build
• Requires GPU; CPU mode exists but is “extremely slow”
• Needs CUDA 7.0, CuDNN 3.0, Python 2.7
• ~1.5GB GPU memory with CuDNN, ~5GB without

Caveats

• Python 2.7 and CUDA 7.0 are long dead; expect archaeology
• The “side-refinement” part from the paper is not implemented in this release
• CPU implementation is explicitly “non-optimal”

Verdict Worth a look if you’re studying historical text detection methods or need to reproduce the ECCV 2016 paper. Skip it if you want a modern, maintained OCR pipeline—this is a research artifact frozen in 2016 dependency hell.