Face detection using bob.ip.facedetect

[Amir Mohammadi]

Hi,

I want to use bob.ip.facedetect but I don't have much experience in face detection.
I have tried using the simple api

bob.ip.facedetect.detect_all_faces(face_image, threshold=20)

but It does not work properly (I have attached the results).

Sorry for asking basic questions but how can I improve this?

Thanks,
Amir

[Tiago Freitas Pereira]

Hey Amir,

Could you please attach the original image?

I have a felling this is related to the quality of the model that was trained.

Another thing that you can try is to play with the sampler object (http://pythonhosted.org/bob.ip.facedetect/py_api.html#bob.ip.facedetect.Sampler) and provide a higher lowest scale. Since your task is face detection with mobile phones, I guess you can use bigger patches to start the face search. Just a feeling, I'm not an expert in face detection.

CHeers

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+unsubscribe@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/
---
You received this message because you are subscribed to the Google Groups "bob-devel" group.
To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Tiago

[Amir Mohammadi]

Hi Tiago,

Here is the original photo.

Thanks,

Amir

On Mon, Sep 19, 2016 at 2:03 PM Tiago Freitas Pereira <tiagofr...@gmail.com> wrote:

Hey Amir,

Could you please attach the original image?

I have a felling this is related to the quality of the model that was trained.

Another thing that you can try is to play with the sampler object (http://pythonhosted.org/bob.ip.facedetect/py_api.html#bob.ip.facedetect.Sampler) and provide a higher lowest scale. Since your task is face detection with mobile phones, I guess you can use bigger patches to start the face search. Just a feeling, I'm not an expert in face detection.

CHeers

On Mon, Sep 19, 2016 at 1:36 PM, Amir Mohammadi <183....@gmail.com> wrote:

Hi,

I want to use bob.ip.facedetect but I don't have much experience in face detection.
I have tried using the simple api

bob.ip.facedetect.detect_all_faces(face_image, threshold=20)

but It does not work properly (I have attached the results).

Sorry for asking basic questions but how can I improve this?

Thanks,
Amir

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+...@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+...@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--

Tiago

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+...@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+...@googlegroups.com.

[Manuel Günther]

Tiago, Amir,

the first idea would be to take an image that shows an upright face. Your image is laying on the side.

The model (cascade) that is used by default was trained partially on the MOBIO training set. Hence, it should be a bit less sensitive to the illumination in MOBIO images. See here: http://pythonhosted.org/bob.ip.facedetect/guide.html#the-shipped-cascade for a complete list of databases used for training.

However, the face detection model is mainly trained on frontal faces, while your test face is shot from below. It should not matter too much, but this might be a reason for mis-detection.

As Tiago said, you can try out several options to improve the results. For example, you might adapt the Sampler to search only for large faces (by increasing the ``lowest_scale``), or you can lower your threshold, so that more possible bounding boxes are returned.

Cheers

Manuel

[Tiago Freitas Pereira]

Hey Amir,

Are you not rotating the image?

In my first test rotating the image in -90 degrees I got this

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+unsubscribe@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+unsubscribe@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--

Tiago

[Amir Mohammadi]

I was plotting the image incorrectly using np.swapaxes instead of np.moveaxes and that was transposing the image too so I thought the image is upright when loaded.

Thanks,
Amir

On Tue, Sep 20, 2016, 10:02 AM Tiago Freitas Pereira <tiagofr...@gmail.com> wrote:

Hey Amir,

Are you not rotating the image?

In my first test rotating the image in -90 degrees I got this

On Tue, Sep 20, 2016 at 3:16 AM, 'Manuel Günther' via bob-devel <bob-...@googlegroups.com> wrote:

Tiago, Amir,

the first idea would be to take an image that shows an upright face. Your image is laying on the side.

The model (cascade) that is used by default was trained partially on the MOBIO training set. Hence, it should be a bit less sensitive to the illumination in MOBIO images. See here: http://pythonhosted.org/bob.ip.facedetect/guide.html#the-shipped-cascade for a complete list of databases used for training.

However, the face detection model is mainly trained on frontal faces, while your test face is shot from below. It should not matter too much, but this might be a reason for mis-detection.

As Tiago said, you can try out several options to improve the results. For example, you might adapt the Sampler to search only for large faces (by increasing the ``lowest_scale``), or you can lower your threshold, so that more possible bounding boxes are returned.

Cheers

Manuel

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+...@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+...@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--

Tiago

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+...@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+...@googlegroups.com.

[Tiago Freitas Pereira]

Ok, cool.

It means that our detector is working fine :-)

Cheers

On Tue, Sep 20, 2016 at 10:05 AM, Amir Mohammadi <183....@gmail.com> wrote:

I was plotting the image incorrectly using np.swapaxes instead of np.moveaxes and that was transposing the image too so I thought the image is upright when loaded.

Thanks,
Amir

On Tue, Sep 20, 2016, 10:02 AM Tiago Freitas Pereira <tiagofr...@gmail.com> wrote:

Hey Amir,

Are you not rotating the image?

In my first test rotating the image in -90 degrees I got this

On Tue, Sep 20, 2016 at 3:16 AM, 'Manuel Günther' via bob-devel <bob-...@googlegroups.com> wrote:

Tiago, Amir,

the first idea would be to take an image that shows an upright face. Your image is laying on the side.

The model (cascade) that is used by default was trained partially on the MOBIO training set. Hence, it should be a bit less sensitive to the illumination in MOBIO images. See here: http://pythonhosted.org/bob.ip.facedetect/guide.html#the-shipped-cascade for a complete list of databases used for training.

However, the face detection model is mainly trained on frontal faces, while your test face is shot from below. It should not matter too much, but this might be a reason for mis-detection.

As Tiago said, you can try out several options to improve the results. For example, you might adapt the Sampler to search only for large faces (by increasing the ``lowest_scale``), or you can lower your threshold, so that more possible bounding boxes are returned.

Cheers

Manuel

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+unsubscribe@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+unsubscribe@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--

Tiago

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+unsubscribe@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+unsubscribe@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--
-- You received this message because you are subscribed to the Google Groups bob-devel group. To post to this group, send email to bob-...@googlegroups.com. To unsubscribe from this group, send email to bob-devel+unsubscribe@googlegroups.com. For more options, visit this group at https://groups.google.com/d/forum/bob-devel or directly the project website at http://idiap.github.com/bob/

---
You received this message because you are subscribed to the Google Groups "bob-devel" group.

To unsubscribe from this group and stop receiving emails from it, send an email to bob-devel+unsubscribe@googlegroups.com.

For more options, visit https://groups.google.com/d/optout.

--

Tiago