| 1. The captured scenes include: carving, boots, baskets, cinema, chef. |
| The simulated scenes include lego, chair, ficus, hotdog, bench. |
| These have different structures and loaders. |
|
|
| In the root of the dataset folder you can also find the `intrinsics.npy` which is the set of captured rays and parameters used in training and `pulse_low_flux.mat` which is the calibrated laser pulse. |
|
|
|
|
| 2. Training transform files. |
|
|
| For the *captured* scenes the training transforms are all named `transforms_train.json` and can be found under `<scene_name>/final_cams/<num_views>/transforms_train.json`. |
|
|
| For the *simulated* scenes the training transforms are all named `transforms_train_v{i}.json` where `i` is the number of views (2, 3, 5). and can be found under `<scene_name>/transforms_train_v{i}.json`. |
|
|
|
|
| 3. Test transform files. |
|
|
| For the *captured* scenes the test transforms can be found under `<scene_name>/final_cams/test_jsons/transforms_test.json`. |
|
|
| For the *simulated* scenes the test transforms can be found under `<scene_name>/transforms_test_final.json`. |
|
|
|
|
| 4. Downloading. |
|
|
| - You can download the data yourself either through the Dropbox download button, or by right-clicking a folder and selecting copy link address, then |
| ``` |
| wget "copied link" |
| ``` |
| will start a download of the folder. |
|
|
| 5. (!!!) Using the dataset. |
|
|
| To use the dataset alongside its transforms please look at the loader in `loaders/loader_captured.py` in the [GitHub repository](https://github.com/anaghmalik/TransientNeRF). Most importantly to use the temporal *captured* data, you will have to resample the transient using the shift given in the `intrinsics.npy` file: |
|
|
| ``` |
| img_shape = (512, 512) |
| exposure_time= 299792458*4e-12 |
| n_bins = 1500 |
| |
| x = (torch.arange(img_shape[0], device="cpu")-img_shape[0]//2+0.5)/(img_shape[0]//2-0.5) |
| y = (torch.arange(img_shape[0], device="cpu")-img_shape[0]//2+0.5)/(img_shape[0]//2-0.5) |
| z = torch.arange(n_bins*2, device="cpu").float() |
| X, Y, Z = torch.meshgrid(x, y, z, indexing="xy") |
| Z = Z*exposure_time/2 |
| Z = Z - shift[0] |
| Z = Z*2/exposure_time |
| Z = (Z-n_bins*2//2+0.5)/(n_bins*2//2-0.5) |
| grid = torch.stack((Z, X, Y), dim=-1)[None, ...] |
| del X |
| del Y |
| del Z |
| |
| rgb = torch.Tensor(rgb)[..., :3000].float().cpu() |
| rgb = torch.nn.functional.grid_sample(rgb[None, None, ...], grid, align_corners=True).squeeze().cpu() |
| |
| ``` |
|
|
| where `shift` is the value from the `intrinsics.npy` file and `rgb` is the original loaded transient. |
|
|
