A TensorFlow Faster R-CNN implementation that sends you away

What it does

This repository contains a TensorFlow reimplementation of the Faster R-CNN object detection framework, originally ported from the Caffe-based py-faster-rcnn. It trains and runs region-based convolutional networks on PASCAL VOC and MS COCO using backbones like VGG16, ResNet, and MobileNet.

The interesting bit

What makes it unusual is its honesty: the README opens by declaring the project deprecated and actively points readers to TensorPack’s more modern implementation. The author also published a technical report documenting minor modifications—such as using crop_and_resize, excluding ground-truth boxes from RoIs during training, and keeping small proposals—which reportedly give slight improvements over the original baseline.

Key highlights

• Supports VGG16, ResNet V1, and MobileNet V1 backbones with released pre-trained models.
• Includes built-in TensorBoard visualization for ground truth, predictions, loss statistics, and activations.
• Supports resuming training with stored numpy random states and image permutations.
• Implements a joint train-and-validation loop to monitor overfitting without reloading the model.
• Reports baseline mAP numbers on VOC 2007 and COCO 2014 (e.g., 79.8 mAP on VOC 07+12 with ResNet101).

Caveats

• Explicitly deprecated by the author, who recommends TensorPack for modern faster/mask R-CNN work.
• Targets very old TensorFlow versions (r1.2 format, with legacy branches for r0.12 and r1.0) and CUDA 8.
• Training cannot be made fully deterministic even with fixed random seeds, and the provided Docker image is no longer valid.

Verdict

Worth a look if you are studying the evolution of detection frameworks or need the specific ablation details from the accompanying technical report. Avoid it if you actually need to train or deploy a modern object detector today.