hikapok/x-detector

Light-Head RCNN and One Novel Object Detector

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

x-detector's Issues

can you give a pbtxt for you light-head-rcnn?

i want to convert you light-head rcnn ckpt to pb,but i don't know the output_name of graph.
can you share the pbtxt for for your light-head-rcnn ckpt?

I want to train/test xception model with light_head_rfcn_train.py, can you Provide a simple script？

I want to train/test xception model with light_head_rfcn_train.py, can you Provide a simple script？
thank you very much

Have you trained the original light_head_rcnn？

I have trained light_head_rcnn on VOC and i find the map is only about 20%.

how to test with checkpoint files?

i want to test on image with checkpoint files you provide， can you provide a simple script for me. thank you very much!

compilation error while building PSROIPooling

Hi all,
I can't build libps_roi_align.so, running into following error:
X-Detector/cpp/PSROIPooling$ cmake .
X-Detector/cpp/PSROIPooling$ make
[ 10%] Building NVCC (Device) object CMakeFiles/cuda_compile_4.dir/cuda_compile_4_generated_rotated_ps_roi_align_grad_op.cu.o
sh: 1: cannot create -/rotated_ps_roi_align_grad_op.cpp1.ii: Directory nonexistent
CMake Error at cuda_compile_4_generated_rotated_ps_roi_align_grad_op.cu.o.cmake:219 (message):
Error generating
/data/open_source_projects/object_detection/X-Detector/cpp/PSROIPooling/CMakeFiles/cuda_compile_4.dir//./cuda_compile_4_generated_rotated_ps_roi_align_grad_op.cu.o

CMakeFiles/ps_roi_align.dir/build.make:84: recipe for target 'CMakeFiles/cuda_compile_4.dir/cuda_compile_4_generated_rotated_ps_roi_align_grad_op.cu.o' failed
make[2]: *** [CMakeFiles/cuda_compile_4.dir/cuda_compile_4_generated_rotated_ps_roi_align_grad_op.cu.o] Error 1
CMakeFiles/Makefile2:67: recipe for target 'CMakeFiles/ps_roi_align.dir/all' failed
make[1]: *** [CMakeFiles/ps_roi_align.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2

Env: OS Ubuntu 18, python 3.6.6, tensorflow-gpu==1.11.0, cuda-9.0

Any help is highly appreciated

Some questions about 74.85%mAP

I want to know what the backbone and the input size are on this map.

what about the speed and mAP vs yolov3

@HiKapok hi, have you ever try to train on COCO ? can the model be as good as the paper result (we can find its' mAP is close to yolov3, but 3 times fps to yolov2). Recently, I want to do human head detection, i want to use light-head-rcnn, can you give some advices? I am new in deep learning,thank you.

Questions about the light_head_rfcn_train.py

I'm trying to run the light_head_rfcn_train.py but I get the error below. It shows that package 'tf_extended' doesn't exist.

Error during loading lips_roi_align.so

Hi, thanks for sharing the code!
I went to cpp/PSROIPooling, and then executed cmake, and finally make, to generate the libps_roi_align.so file. But when I run the demo, I encountered this error:

tensorflow.python.framework.errors_impl.NotFoundError: ./cpp/PSROIPooling/libps_roi_align.so: undefined symbol: _ZN10tensorflow11GetNodeAttrERKNS_9AttrSliceENS_11StringPieceEPSs

I used exactly the same environment: tensorflow-1.6, python-3.5 and ubuntu 16.04. Also, I googled to find relevant topics but with no success.
Appreciate it if you can help.

Why tf tries to occupy so much GPU memory for simple test?

Hi there,
I'm trying to run your light_head_simple_demo.py with provided model model.ckpt-121397
But while starting the script I run into out of memory issue: while restoring parameters from the checkpoint tf tries to occupy > 14Gb of GPU and falls because I have less than 16Gb.
But the model weight file has size less than 0,5Gb.
Does it really require >16Gb? Or I have somewhere wrong settings for the model? How much memory does it requires for that test?

For test I use tensorflow docker container with pre-installed Ubuntu16, cuda9.0, tensorflow1.6, python 3.5

Thanks

cpp psroi pooling can not built

Same error:

oi_align_op.h:25:61: fatal error: third_party/eigen3/unsupported/Eigen/CXX11/Tensor: No such file or directory
compilation terminated.
CMake Error at cuda_compile_4_generated_rotated_ps_roi_align_grad_op.cu.o.cmake:220 (message):
  Error generating
  /media/jintain/sg/ai/career/robot/ctirobot/detect/X-Detector/cpp/PSROIPooling/build/CMakeFiles/cuda_compile_4.dir//./cuda_compile_4_generated_rotated_ps_roi_align_grad_op.cu.o


CMakeFiles/ps_roi_align.dir/build.make:84: recipe for target 'CMakeFiles/cuda_compile_4.dir/cuda_compile_4_generated_rotated_ps_roi_align_grad_op.cu.o' failed
make[2]: *** [CMakeFiles/cuda_compile_4.dir/cuda_compile_4_generated_rotated_ps_roi_align_grad_op.cu.o] Error 1
CMakeFiles/Makefile2:72: recipe for target 'CMakeFiles/ps_roi_align.dir/all' failed
make[1]: *** [CMakeFiles/ps_roi_align.dir/all] Error 2
Makefile:83: recipe for target 'all' failed
make: *** [all] Error 2