What it does

This repository collects pre-trained Caffe models and their .prototxt deploy files for a wide range of classic computer vision architectures: ResNet, ResNeXt, Inception v3/v4, Inception-ResNet, DenseNet, wide ResNet, aligned Inception-ResNe(X)t, DPNs, and others. It covers three standard tasks—classification on ImageNet, object detection with Faster R-CNN on PASCAL VOC, and semantic segmentation with PSPNet on PASCAL VOC—each with published accuracy, speed, and memory numbers.

The interesting bit

The README is admirably honest: “Most of the pre-train models are converted from other projects, the main contribution belongs to the original authors.” This is essentially a curation and conversion effort, but done at a time when converting between MXNet, TensorFlow, Keras, and Caffe weight formats was genuinely tedious work. The tables are the value—systematic head-to-head numbers for model selection in 2016–2017.

Key highlights

• Classification: Inception-ResNet-v2 hits 19.86% top-1 / 4.83% top-5 error (single-crop 299×299); ResNet-269-v2 reaches 17.87% / 3.85% with 12-crop evaluation
• Detection: Inception-v4 leads the Faster R-CNN table at 81.49 mAP@50 on VOC 2007 test
• Segmentation: ResNet-101-v2 achieves 77.94% mIoU with PSPNet on VOC 2012 val
• Explicit speed and memory metrics throughout (e.g., ResNeXt-101-64×4d needs 11,277 MB training memory at batch=96)
• MIT licensed; designed to plug into the companion py-RFCN-priv repo for evaluation and fine-tuning

Caveats

• Caffe itself is effectively archived; these models are historical artifacts unless you’re maintaining legacy pipelines
• Some table cells are incomplete (e.g., DPN-107 12-crop results show “../..”)
• The segmentation table omits speed/memory for Inception-v4 entirely

Verdict

Worth bookmarking if you’re resurrecting an old Caffe codebase, writing a history-of-ImageNet-models post, or need exact .caffemodel weights for reproducibility. Everyone else has moved to PyTorch Hub and timm.