Fangneng Zhan

Besides, I also test the sample in examples/vae_reference.py, still find Gumbel outperform reinforce and rebar. (n_samples=10, final loss: 122 for Gumbel, 157 for reinforce, 152 for rebar).

> Hi @HEmile Thanks for your reply. In my test, the latent space size in examples/vae/discrete_vae.py and examples/vae_reference.py are 20×10 and 2×10 (i.e., the default setting), respectively. I run the...

Thanks for sharing the information and your awesome dataset. We find there is a typo when retrieving the information from Table 1 of 'Deep Image Synthesis from Intuitive User Input:...

A generation network is followed to translate gaussian map to illumination maps, i.e., GenProjector. The generation network is highly biased to the training scene, I recommend you to train it...

You can remove the depth part, which doesn't bring clear improvement. I got the depth map by contacting the author of Laval indoor dataset.

Hi, they are computed on the images rendered with illumination maps. For angular error, please refer to DeepLight: Learning Illumination for Unconstrained Mobile Mixed Reality

I evaluate the results with the pretrained DRN model, and find the mAP using drn-d-22-cityscapes.pth is only 21, but the mAP using drn-d-105-ms-cityscapes.pth can match the reported mAP in the...

Hi, do you solve it? I train the task of horse2zebra with the default setting and get a FID of 61.68, which is much lower than the reported 45.5. Is...

Seems the best score is not in the final epoch, this work may report the best score during training. You can check this https://github.com/lyndonzheng/F-LSeSim/issues/1

I include the depth part in a previous version, but remove it later as observing most users don't have access to the depth of Laval indoor dataset.