Docs

docs/source/examples/ngp.rst

@@ -14,17 +14,17 @@ memory footprint is about 3GB.
 .. note::
 
     The paper makes use of the alpha channel in the images to apply random background
-    augmentation during training. Yet we only uses a constant white background.
+    augmentation during training. Yet we only uses RGB values with a constant white background.
 
-+----------------------+-------+-------+------------+-------+--------+--------+--------+--------+
-|                      | Lego  | Mic   | Materials  |Chair  |Hotdog  | Ficus  | Drums  | Ship   |
-|                      |       |       |            |       |        |        |        |        |
-+======================+=======+=======+============+=======+========+========+========+========+
-| Paper (PSNR: 5min)   | 36.39 | 36.22 | 29.78      | 35.00 | 37.40  | 33.51  | 26.02  | 31.10  |
-+----------------------+-------+-------+------------+-------+--------+--------+--------+--------+
-| Ours  (PSNR)         | 36.61 | 37.45 | 30.15      | 36.10 | 37.88  | 32.07  | 25.83  |        |
-+----------------------+-------+-------+------------+-------+--------+--------+--------+--------+
-| Ours  (Training time)| 300s  | 272s  | 258s       | 311s  | 275s   | 254s   | 249s   |        |
-+----------------------+-------+-------+------------+-------+--------+--------+--------+--------+
++----------------------+-------+-------+------------+-------+--------+--------+--------+--------+--------+
+|                      | Lego  | Mic   | Materials  |Chair  |Hotdog  | Ficus  | Drums  | Ship   | AVG    |
+|                      |       |       |            |       |        |        |        |        |        |
++======================+=======+=======+============+=======+========+========+========+========+========+
+| Paper (PSNR: 5min)   | 36.39 | 36.22 | 29.78      | 35.00 | 37.40  | 33.51  | 26.02  | 31.10  | 33.18  |
++----------------------+-------+-------+------------+-------+--------+--------+--------+--------+--------+
+| Ours  (PSNR:4.5min)  | 36.71 | 36.78 | 29.06      | 36.10 | 37.88  | 32.07  | 25.83  | 31.39  | 33.23  |
++----------------------+-------+-------+------------+-------+--------+--------+--------+--------+--------+
+| Ours  (Training time)| 286s  | 251s  | 250s       | 311s  | 275s   | 254s   | 249s   | 255s   | 266s   |
++----------------------+-------+-------+------------+-------+--------+--------+--------+--------+--------+
 
 .. _`github repository`: : https://github.com/KAIR-BAIR/nerfacc/

docs/source/examples/vanilla.rst

@@ -1,2 +1,21 @@
 Vanilla Nerf 
 ====================
+
+See code `examples/train_mlp_nerf.py` at our `github repository`_ for details.
+
+Benchmarks
+------------
+
+
++----------------------+-------+-------+------------+-------+--------+--------+--------+--------+
+|                      | Lego  | Mic   | Materials  |Chair  |Hotdog  | Ficus  | Drums  | Ship   |
+|                      |       |       |            |       |        |        |        |        |
++======================+=======+=======+============+=======+========+========+========+========+
+| Paper (PSNR: 5min)   | 32.54 | 32.91 | 29.62      | 33.00 | 36.18  | 30.13  | 25.01  | 28.65  |
++----------------------+-------+-------+------------+-------+--------+--------+--------+--------+
+| Ours  (PSNR)         | XX.XX | XX.XX | XX.XX      | XX.XX | XX.XX  | XX.XX  | XX.XX  | XX.XX  |
++----------------------+-------+-------+------------+-------+--------+--------+--------+--------+
+| Ours  (Training time)| XXmin | XXmin | XXmin      | 45min | XXmin  | XXmin  | 41min  | XXmin  |
++----------------------+-------+-------+------------+-------+--------+--------+--------+--------+
+
+.. _`github repository`: : https://github.com/KAIR-BAIR/nerfacc/

docs/source/index.rst

@@ -7,8 +7,8 @@ Using NerfAcc,
 
 - The `vanilla Nerf model`_ with 8-layer MLPs can be trained to *better quality* (+~1.0 PNSR) \
   in *45 minutes* rather than *1~2 days* as in the paper.
-- The `instant-ngp Nerf model`_ can be trained to *better quality* (+~1.0 PNSR) \
-  in *5 minutes* compare to the paper.
+- The `instant-ngp Nerf model`_ can be trained to *equal quality* with *9/10th* of the training time \
+  comparing to the official pure-CUDA implementation.
 - The `D-Nerf model`_ for *dynamic* objects can also be trained in *45 minutes* \
   rather than *2 days* as in the paper, and with *better quality* (+~2.0 PSNR).
 - *Unbounded scenes* from `MipNerf360`_ can also be trained in \

examples/train_mlp_dnerf.py

@@ -72,16 +72,20 @@
     ).item()
 
     # setup the radiance field we want to train.
-    max_steps = 40000
+    max_steps = 50000
     grad_scaler = torch.cuda.amp.GradScaler(1)
     radiance_field = DNeRFRadianceField().to(device)
     optimizer = torch.optim.Adam(radiance_field.parameters(), lr=5e-4)
     scheduler = torch.optim.lr_scheduler.MultiStepLR(
         optimizer,
-        milestones=[max_steps // 2, max_steps * 3 // 4, max_steps * 9 // 10],
+        milestones=[
+            max_steps // 2,
+            max_steps * 3 // 4,
+            max_steps * 5 // 6,
+            max_steps * 9 // 10,
+        ],
         gamma=0.33,
     )
-
     # setup the dataset
     data_root_fp = "/home/ruilongli/data/dnerf/"
     target_sample_batch_size = 1 << 16

examples/train_mlp_nerf.py

@@ -87,13 +87,18 @@
         ).item()
 
     # setup the radiance field we want to train.
-    max_steps = 40000
+    max_steps = 50000
     grad_scaler = torch.cuda.amp.GradScaler(1)
     radiance_field = VanillaNeRFRadianceField().to(device)
     optimizer = torch.optim.Adam(radiance_field.parameters(), lr=5e-4)
     scheduler = torch.optim.lr_scheduler.MultiStepLR(
         optimizer,
-        milestones=[max_steps // 2, max_steps * 3 // 4, max_steps * 9 // 10],
+        milestones=[
+            max_steps // 2,
+            max_steps * 3 // 4,
+            max_steps * 5 // 6,
+            max_steps * 9 // 10,
+        ],
         gamma=0.33,
     )