search

DLR-RM / AugmentedAutoencoder

Official Code: Implicit 3D Orientation Learning for 6D Object Detection from RGB Images
MIT License
338 stars 97 forks source link

an issue to train autoencoder with LINEMOD 3D models #29

Closed mengce97 closed 5 years ago

mengce97 commented 5 years ago

when i train an autoencoder with linemod datasets,i change the cfg like this:

[Dataset] MODEL: reconst H: 128 W: 128 C: 3 RADIUS: 500 RENDER_DIMS: (650, 484) K: [572.4114, 0, 325.2611, 0, 573.5704, 242.0489, 0, 0, 1]

Scale vertices to mm

VERTEX_SCALE: 10 ANTIALIASING: 8 PAD_FACTOR: 1.2 CLIP_NEAR: 10 CLIP_FAR: 10000 NOOF_TRAINING_IMGS: 20000 NOOF_BG_IMGS: 15000

The vertex scale of linemod is cm, so i change the vertex_scale, the inner cammera matrix of linemod is K: [572.4114, 0, 325.2611, 0, 573.5704, 242.0489, 0, 0, 1], and so on as the render_dims. However, i still encounter the problem that "object in rendering not visible,have you scaled the vertices to mm?". Are there any mistakes with the parameters i changed? Any parameters else i need to change? Thanks for your reply!

mengce97 commented 5 years ago

oh,i seem to use the wrong 3D model,however,can u help me with the rest questions? Is the K and render_dims correct?

MartinSmeyer commented 5 years ago

K seems correct, the render dimensions should be 640x480. Vertex scale = 1 since cm is already converted to mm in the code. For the larger objects, you might have to increase the radius to 800 to ensure everything is in the image..

mengce97 commented 5 years ago

thank u!