Improving the framework for 2D detection

31 views
Skip to first unread message

Gautam Malu

unread,
Apr 2, 2017, 3:30:55 PM4/2/17
to opendetection
Hi,

I am a graduate student working in computer-vision at Internation Institute of Information Technology Hyderabad. I am applying for GSOC 2017 for this library. I want to work on CNNs for object detection idea. I have installed the library and tested out examples.
I have prepared a brief report on current state of art CNN methods for 2D object detections  https://github.com/gautamMalu/opendetection/blob/algo_report/ObjectDetection.md

I have following queries regarding this idea:
On the project page, it is mentioned that
 
  • Implement OD interface for training and testing (under od::ODTrainer s) neural network architectures with convolutional layer and possible recurrent layers using Caffe.


Is there a particular reason for using Caffe, I am asking because Tensorflow also have C++ API, and it's better maintained than Caffe. 



  • Add to OpenDitection the following algorithms: .. 1. Classification - Types: different options - AlexNet, VGG, Resnet50, Resnet150, GoogLenet. .. 2. Detection - Faster RCNN .. 3. Location probability map

OD is for object detection so why are focusing on Classification instead of detection?

Shyam Nandan Rai

unread,
Apr 5, 2017, 6:12:02 AM4/5/17
to opendetection
Hi Gautam,

I think the main idea of implementing classification is use of probability location map.As this method localizes the object using classification (though we use last convlution layer for this method)

For more details please refer the link : https://arxiv.org/abs/1311.2901

Gautam Malu

unread,
Apr 5, 2017, 6:38:26 AM4/5/17
to Shyam Nandan Rai, opendetection, kripasind...@dfki.de
Hi Shyam,

Thanks for the clarification. If that is the idea behind the classification, we can use these state of the art gradient based visualization techniques:

http://cnnlocalization.csail.mit.edu/Zhou_Learning_Deep_Features_CVPR_2016_paper.pdf
https://arxiv.org/abs/1610.02391

You can find the implementation of 2nd paper for VGG16 and ResNet50 here (https://github.com/gautamMalu/keras-grad-cam)
However, I am still not so clear why we are using these unsupervised localization techniques instead of state of the art supervised techniques such as faster-rcnn, YOLO. I guess data preparation for the training of these techniques is little challenging but we can start with providing the pre-trained models as detectors first and then figure out the training of these models also.


Best regards,
Gautam Malu


--
You received this message because you are subscribed to a topic in the Google Groups "opendetection" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/opendetection/wQb4VZOgI2c/unsubscribe.
To unsubscribe from this group and all its topics, send an email to opendetection+unsubscribe@googlegroups.com.
To post to this group, send email to opende...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/opendetection/d2f6e169-aced-4761-a1f1-9c70e9b187a9%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Shyam Nandan Rai

unread,
Apr 5, 2017, 7:04:19 AM4/5/17
to opendetection
Hi Gautam,

Yes, you are right. If we start training for faster r-cnn it takes high computation resources (at least 3gb for training on small dataset). So, it would be better to start with pre- trained models.

Gautam Malu

unread,
Apr 17, 2017, 2:29:18 PM4/17/17
to Shyam Nandan Rai, kripasind...@dfki.de, opendetection
Hi,
I have implemented grad-CAM visualization technique (https://arxiv.org/abs/1610.02391) for Caffe too https://github.com/gautamMalu/caffe-gradCAM

Best regards,
Gautam Malu

On Wed, Apr 5, 2017 at 4:34 PM, Shyam Nandan Rai <shya...@iiits.in> wrote:
Hi Gautam,

Yes, you are right. If we start training for faster r-cnn it takes high computation resources (at least 3gb for training on small dataset). So, it would be better to start with pre- trained models.
--
You received this message because you are subscribed to a topic in the Google Groups "opendetection" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/opendetection/wQb4VZOgI2c/unsubscribe.
To unsubscribe from this group and all its topics, send an email to opendetection+unsubscribe@googlegroups.com.
To post to this group, send email to opende...@googlegroups.com.
Reply all
Reply to author
Forward
0 new messages