/X-Detector

Light-Head RCNN and One Novel Object Detector

Primary LanguagePythonApache License 2.0Apache-2.0

X-Detector

X-Detector is a collection of several object detection algorithms. And some of those have not appeared in any academic papers.

Up to now, this repository contains code of the re-implementent of Light-Head R-CNN and this trained version had got to 74.85%mAP(77.29%mAP using VOC12 evaluation alogorithm), more improvement is still in process. You can download the latest model weights (trained on PASCAL-07+12) from GoogleDrive.

Below is the training timeline of Light-Head RCNN for single 480x480 input image.

Besides, several other detectors(named "X-Det") are also included, the main idea behind "X-Det" is to introduce explicit attention mechanisms between feature map channels. But the current performance of them is only ~0.71mAP on PASCAL-VOC 2007 Test Dataset, more improvement may need to introduce FPN-like structure on the top feature map(I didn't try this which is beyond the initial purpose of "X-Det").

The pre-trained weights of backbone network can be found in Resnet-50 backbone and Xception backbone. The latest version of PsRoIAlign is here.

You can use part of these codes for your research purpose, without any permission but following Apache License 2.0. All codes were tested under TensorFlow 1.6, Python 3.5, Ubuntu 16.04.

Here are some demo result images of "X-Det"-V2, debugging is still in process to make better performance:

Update:

  • More than 7x performance improvement for Light-Head RCNN.
  • Fine-tunning modified resnet backbone for "X-Det".
  • Reorganize the order of the preprocessing pipeline.
  • Switch to sample-wise hard negtive mining.

Apache License 2.0