Skip to content

Latest commit

 

History

History
176 lines (171 loc) · 6.07 KB

pretrained_models.md

File metadata and controls

176 lines (171 loc) · 6.07 KB
Raw

Plan2Scene Models

Trained Plan2Scene models are shown below.

Texture Synthesis Model Texture Propagation Model Observed Surfaces Unobserved Surfaces All Surfaces
Color Freq Subs Subs**
(Version2)		 FID Tile Color Freq Subs Subs**
(Version2)		 FID Tile Color Freq Subs Subs**
(Version2)		 FID Tile
Synth (ours)
CVPR Version		 [texture_synth_conf.yml] [texture_prop_conf.json] 0.431 0.035 0.350 0.463 196.1 16.4 0.653 0.032 0.393 0.490 199.4 18.6 0.591 0.034 0.392 0.485 196.2 17.6
Synth (ours)
Version 2*		 [texture_synth_conf.yml] [Weights] [texture_prop_conf.json] [Weights] 0.386 0.027 0.382 0.480 158.8 11.0 0.714 0.028 0.413 0.461 178.9 12.8 0.579 0.028 0.380 0.480 166.9 12.4
* Synth version 2 is trained on the stationary textures dataset version 2.
** Subs metric version 2 is trained on the stationary textures dataset version 2 and the open-surfaces dataset.

To use a pre-trained model,

  1. Download the texture_synth_conf.yml file, texture synthesis model weights file, texture_prop_conf.json file and texture propagation model weights file.
  2. Update ./conf/plan2scene/texture_gen.json file as follows.
    • For the field texture_synth_conf, specify the path to the texture_synth_conf.yml file downloaded.
    • For the field checkpoint_path, specify the path to the texture synthesis model weights file downloaded.
  3. When required, specify the command line arguments as follows. (E.g. In the gnn_texture_prop.py script.)
    • For the GNN_PROP_CONF_PATH argument, specify the path to the texture_prop_conf.json file.
    • For the GNN_PROP_CHECKPOINT_PATH argument, specify the path to the texture propagation model weights.

Substance Classifier Models

Trained models for the substance classifier of the SUBS metric are shown below. All models are trained following the approach described in the supplementary material section of the paper.

Train Set Validation Set Validation Accuracy Config Weights
Version 1
CVPR Paper
  • Stationary Textures Dataset (4120 Crops)
  • OpenSurfaces Dataset (6378 Crops)
  • Stationary Textures Dataset (64 Crops)
  • OpenSurfaces Dataset (60 Crops)
 87.9% [substance_classifier_conf.json]
Version 2
  • Stationary Textures Dataset Version 2 (6440 Crops)
  • OpenSurfaces Dataset (6378 Crops)
  • Stationary Textures Dataset Version 2 (64 Crops)
  • OpenSurfaces Dataset (60 Crops)
 89.1% [substance_classifier_conf.json] [Weights]

To use a pre-trained model,

  1. Download the substance_classifier_conf.yml file and the weights file.
  2. Update ./conf/plan2scene/metrics.json file as follows.
    • For the field substance_classifier.conf_path, specify the path to the substance_classifier_conf.yml file.
    • For the field substance_classifier.checkpoint_path, specify the path to the weights file.