hikapok/ron_tensorflow

A SOTA RON Detector in TensorFlow

RON Detector in TensorFlow: Reverse Connection with Objectness Prior Networks for Object Detection

RON is an efficient object detection system as descibed in This CVPR paper.

This repository contains code of the re-implement of RON following the above paper. Now almost all of the implementation details matches the open-source version by the the author of RON.

The code is modified from SSD-Tensorflow. You can use the code to train/evaluate your network for object detection task.

For more details (including dataset prepare), please refer to README of SSD-Tensorflow.

Update: Recently, I have found some details of the detection pipeline maybe sub-optimal in this implementation. So if you would like use codes here for further research, I recommend you to refer to this repo which includes many of my latest insights about detection.

Update:

• Add SSD preprocesing method using Tensorflow ops [zero ground truth fixed]
• Modify details of the network to match the original Caffe code
• Add NMS using Tensorflow ops to support two mode
• Fix most part of the matching strategy between ground truth and anchors
• Replica GPU training support (If you are using Tensorflow 1.5.0+, then rename the replicate_model_fn.py)
• Add voc eval with debug
• Add realtime eval, using class-wise bboxes-select and nms
• Add support for training use vgg16_reducedfc model converted from pytorch, you can get from here
• Other important fixes [2018.03.21]
• Make all anchors on different layers be matched together, to avoid some suboptimal matching results
• Refactor anchors matching pipeline
• Fix attribute 'difficult' missing problem in the TFRecords dataset
• Model-320 (reduced version) trained on VOC07+12 dataset now is available at here, the heavier one needs to be trained by yourself(may I will also update a trained model later)

Last Update:

• Maybe the code here is one of the few Tensorflow detection pipelines in open-source that could be trained to get 0.7+mAP(not using weights converted from other framework), except the official object-detection API.
• You can use these code snippets for your own project
• Model trained with heavier VGG-16 backbone got to 0.715mAP(0.734mAP using VOC12 evaluation alogorithm) at 120k steps on my side without other tricks.
• You can try other matching threshold, different learning rate policy, more augumentation, other loss weights to get more improvement. I didn't try these but picked one simple setting.
• If you would like to train using reduced_vgg from here, please make sure that you swaped image channals from the default RGB order into BGR order, and the pixel value should in [-128, 128].

Note: Model trained (07+12 VOC-train and test on VOC07-test) using the initial version of this code can only get to 0.45~0.55mAP, clone the latest version will give you much better performance at 0.7+mAP(needs ~120k steps, training with ron_net.py and evaluation with eval_ron_network.py). Futher improvement is still going on.

Here are some demo result images of reduced-version RON-320 detector(with a heavier vgg16-backbone 0.74mAP is reported in paper) trained using this code:

ron_tensorflow's Issues

Viewing the Objectness prior generated.

Hi there,

Is there a way to view the intermediate objectness prior generated? I believe that it would be great if you could incorporate it.

How to trian with SSD-Net

Hello, HiKapok!
I'm so glad to find out your codes to fix up the loss problem which occurs in SSD-Tensorflow. But I wonder how to train the vgg-SSD-Net by your codes. Should I write a new script like ron_net_multi_gpu.py?

the difference between SSD_tensorflow and your code

Hi,
It confuses me that "Make all anchors on different layers be matched together, to avoid some suboptimal matching results". I cannot figure out what is difference bettween match bbox layer by layer and match bbox together.In fact,we mach every anchor with the gt boxx which has largest overlap with the anchor among all gt bboxes of the input image

trainning question

when I train ron.py file, this mAP is only 55.97%.how to improve this mAP to 0.74? thanks!

where to download the vgg_16.ckpt?

hello,May I ask you where to download the vgg_16.ckpt used in your code? Thanks a lot!

ssd performance reproduction

HI! HiKapok, thanks for your work, can you share your result on ssd after modification compare to the original paper? Thank you.