Description
Hello, great work on the HyperNeRF project!
I’ve been exploring the HyperNeRF dataset and wanted to clarify a few points regarding the resolution of the images. I noticed some discrepancies between the dataset's resolutions and the ones mentioned in the paper. Specifically, the HyperNeRF paper states:
"For our evaluation metrics, we train at half of 1080p resolution (960x540) for 250k iterations, using 128 samples per ray with a batch size of 6,144, which takes roughly 8 hours on 4 TPU v4s. For qualitative results, we train at full-HD (roughly 1920x1080) with 256 samples per ray for 1M iterations, which takes roughly 64 hours."
However, in the dataset I have, I found the following:
1. hypernerf_vrig
- vrig-chicken, vrig-3dprinter, broom:
- 2x Folder: Images are at 536x960 resolution.
- peel-banana:
- 1x Folder: Not available.
- 2x Folder: Images at 536x960 resolution.
2. hypernerf_misc
-
split-cookie, oven-mitts, keyboard, espresso, americano:
- 1x Folder: Images at 1072x1920 resolution.
- 2x Folder: Images resized to 536x960 resolution.
-
cross-hand1:
- 1x Folder: Not available.
- 2x Folder: Images at 536x960 resolution.
-
tamping:
- 1x Folder: Not available.
- 2x Folder: Images at 1080x1920 resolution. (foldering for this scene seems incorrect)
- 4x Folder: Images at 540x960 resolution.
3. hypernerf_interp
- torchocolate, slice-banana, hand1-dense-v2, cut-lemon1, chickchicken, aleks-teapot:
- 2x Folder: Images at 536x960 resolution.
- 1x Folder: Not available.
Looking at the dataset, it seems that except for the tamping dataset (which has 1080x1920 resolution), the original resolution for most images is 1072x1920, and the half resolution is 536x960.
Could you kindly clarify why there are these differences, and if the dataset has been processed in a way that differs from what is described in the paper? Specifically, why is the tamping dataset at 1080x1920 resolution when it should be 960x540 or 536x960 as per the paper?
Thanks in advance for your help, and keep up the amazing work!