random bkgd is useless

README.md

@@ -24,4 +24,5 @@ Tested with the default settings on the Lego test set.
 1. sample rays over all images per iteration (`batch_over_images=True`) is better: `PSNR 33.31 -> 33.75`.
 2. make use of scheduler (`MultiStepLR(optimizer, milestones=[20000, 30000], gamma=0.1)`) to adjust learning rate gives: `PSNR 33.75 -> 34.40`.
 3. increasing chunk size (`chunk: 8192 -> 81920`) during inference gives speedup: `FPS 4.x -> 6.2`
+4. random bkgd color (`color_bkgd_aug="random"`) for the `Lego` scene actually hurts: `PNSR 35.42 -> 34.38`
 

examples/trainval.py

@@ -63,7 +63,7 @@ if __name__ == "__main__":
     train_dataset = SubjectLoader(
         subject_id="lego",
         root_fp="/home/ruilongli/data/nerf_synthetic/",
-        split="train",
+        split="trainval",
         num_rays=8192,
     )
     # train_dataset.images = train_dataset.images.to(device)