owenzlz/instance_insertion

Context-aware Synthesis and Placement of Object Instances

Please find the technique details in the paper

License

Copyright (C) 2018 NVIDIA Corporation. All rights reserved. Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode).

Network Architecture

The network contains two major modules, a "where" module (the first figure) to determine the fesiable location of the object, and a "what" module (the second figure) to generate a proper shape. The two modules are jointly trained, where the blue dashed arrows indicate the linkage of them.

                   

Dataset

• Download and place cityscape dataset at "db_root" in the options.py

How to run the code

• Check options.py and specify your own path accordingly.
• Run main.py, it will save results for pairs of different random vectors, i.e., (z_appr1, z_spatial1), (z_appr2, z_spatial1), and (z_appr1, z_spatial2)

All code tested on Ubuntu 16.04, pytorch 0.3.1, and opencv 3.4.0

Explanation of code details

options.py

• db_root: as explained above
• target_class: person or car
• image_sizex_small: image width when training where module
• image_sizey_small: image height when training where module
• image_sizex_big: image width when training what module
• image_sizey_big: image height when training what module
• compact_sizex: image width of generated object
• compact_sizey: image height of generated object
• embed_dim_small: dim of output of an encoder in where module
• embed_dim_big: dim of output of an encoder in what module

main.py

• Training part starts from line 56

• Between line 56 and 161, it loads training images and check whether it is okay to proceed. We pick 2 seg maps at random. Image 1) b_real_seg_small or b_real_seg_big corresponds to x+ in where and what. It is contains at least one object (variable "has_ins"), then proceed (line 94). Then, check whether there is at least one proper object that are not too small or too narrow (line 120). Image 2) b_cond_seg_small or b_cond_seg_big corresponds to x in where and what. It is just a random image.

• Forward starts at line 161

• Log at line 186

• Save images at line 203

model.py

• Define networks in line 44. Networks are actually defined in networks.py
• Define optimizers in line 114
• Set inputs from line 152-240 We transform a box using A into x+ to prepare real examples, which is done by stn_fix.
• Reparameterize function for VAE in line 241
• Computing edges in line 249-266
• Helper functions in line 268-286
• Forward where supervised in line 288-315
• Forward where/what unsupervised in line 316-374
• Forward what supervised in line 375-399
• Backward for each discriminator in line 401-463
• Backward for generation parts in line 465-539 coord_loss: make sure that the whole compact instance is transformed. stn_theta_loss: preventing to predict too small objects or flipped objects For other losses you can understand what it is by its name.