search

diegoaldarondo / Label3D

Labeling GUI for multi-camera tracking
35 stars 11 forks source link

Support for varying image size across cameras #6

Closed Selmaan closed 4 years ago

Selmaan commented 4 years ago

Label3D currently uses the image size of the first camera (see line 175 in Label3D). Use cases may often have different image sizes for different cameras, and I see no reason why support for this feature would be complicated to implement.

I can modify the code so that image size is recorded independently for each camera, and is then indexed by camera when creating cameraParams arguments (e.g. in line 270 in Label3D)?

spoonsso commented 4 years ago

Interesting, will let Diego address this, but note I will also need to make some adjustments to dannce to accommodate varying image sizes.

Out of curiosity, how are you arriving at different image sizes? Are you using different cameras, or are you cropping differently for each camera?

On Fri, Jun 12, 2020, 5:45 PM Selmaan notifications@github.com wrote:

Label3D currently uses the image size of the first camera (see line 175 in Label3D). Use cases may often have different image sizes for different cameras, and I see no reason why support for this feature would be complicated to implement.

I can modify the code so that image size is recorded independently for each camera, and is then indexed by camera when creating cameraParams arguments (e.g. in line 270 in Label3D)?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/diegoaldarondo/Label3D/issues/6, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAW2P46YH6THF2JFYHMTELLRWKOZDANCNFSM4N4VK5DQ .

spoonsso commented 4 years ago

Different camera make/models, I mean to say

On Fri, Jun 12, 2020, 5:49 PM Tim Dunn dunn.tw@gmail.com wrote:

Interesting, will let Diego address this, but note I will also need to make some adjustments to dannce to accommodate varying image sizes.

Out of curiosity, how are you arriving at different image sizes? Are you using different cameras, or are you cropping differently for each camera?

On Fri, Jun 12, 2020, 5:45 PM Selmaan notifications@github.com wrote:

Label3D currently uses the image size of the first camera (see line 175 in Label3D). Use cases may often have different image sizes for different cameras, and I see no reason why support for this feature would be complicated to implement.

I can modify the code so that image size is recorded independently for each camera, and is then indexed by camera when creating cameraParams arguments (e.g. in line 270 in Label3D)?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/diegoaldarondo/Label3D/issues/6, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAW2P46YH6THF2JFYHMTELLRWKOZDANCNFSM4N4VK5DQ .

Selmaan commented 4 years ago

I'm using identical cameras and lenses, but two of the views require larger image sizes to cover the entire arena.

I have 4 cameras at 2816x1408, and 2 cameras at 2816x1696, high resolution is because it's a small chickadee moving around a big arena. What I've been doing is using a first-pass network on heavily downsampled image to localize the bird, and then cropping data from all cameras to 320x320 centered on the bird. So without knowing too much, I'd guess that only the COM-finder part of dannce needs to handle this, and I think I'm going to need to create my own COM-finder anyway? I don't know enough about the dannce pipeline yet to know what that implies on your end though

Selmaan commented 4 years ago

(to be specific, the data collected is different sizes between cameras because of hardware settings I'm using on the camera hardware, not any post-hoc cropping)

diegoaldarondo commented 4 years ago

Hey Selmaan,

I'd be happy to roll that in. If you want to take a stab at it, feel free to make a branch.

Otherwise you could send me some images, camera parameters, and a skeleton, and I could figure it out.

davidhildebrand commented 4 years ago

Support for different image sizes in Label3D and DANNCE will also be useful for me. My primary reason for cropping at the camera level is to eliminate unnecessary transfer over USB, which can easily saturate with 5 cameras running at or above 100 Hz.

Selmaan commented 4 years ago

@diegoaldarondo is obj.ImageSize used in saving/loading anywhere? I pushed a commit to the multi-image-size branch which is working fine for my use case, but I haven't tested all possible functions. (Basically stepped through all cases where the property was called in Label3D.m, but didn't search elsewhere)

diegoaldarondo commented 4 years ago

Looked over the code. I think it should be fine with the rest. I'll merge after I get a chance to test on the demo.