1 - add configs to reproduce MipNerf360 results

2 - add instructions to build the code on ubuntu
3- add video and names

README.md

@@ -1,5 +1,41 @@
 # 3D Gaussian Splatting as Markov Chain Monte Carlo
 
+[![button](https://img.shields.io/badge/Project%20Website-orange?style=for-the-badge)](https://ubc-vision.github.io/3dgs-mcmc/)
+[![button](https://img.shields.io/badge/Paper-blue?style=for-the-badge)](https://arxiv.org/abs/2404.09591)
+
+<span class="author-block">
+  <a href="https://shakibakh.github.io/">Shakiba Kheradmand</a>,
+</span>
+<span class="author-block">
+  <a href="http://drebain.com/"> Daniel Rebain</a>,
+</span>
+<span class="author-block">
+  <a href="https://hippogriff.github.io/"> Gopal Sharma</a>,
+</span>
+<span class="author-block">
+  <a href="https://wsunid.github.io/"> Weiwei Sun</a>,
+</span>
+<span class="author-block">
+  <a href="https://scholar.google.com/citations?user=1iJfq7YAAAAJ&hl=en"> Yang-Che Tseng</a>,
+</span>
+<span class="author-block">
+  <a href="http://www.hossamisack.com/">Hossam Isack</a>,
+</span>
+<span class="author-block">
+  <a href="https://abhishekkar.info/">Abhishek Kar</a>,
+</span>
+<span class="author-block">
+  <a href="https://taiya.github.io/">Andrea Tagliasacchi</a>,
+</span>
+<span class="author-block">
+  <a href="https://www.cs.ubc.ca/~kmyi/">Kwang Moo Yi</a>
+</span>
+
+<hr>
+
+<video controls>
+  <source src="docs/resources/training_rand_compare/bicycle_both-rand.mp4" type="video/mp4">
+</video>
 
 <section class="section" id="BibTeX">
   <div class="container is-max-desktop content">
@@ -15,12 +51,42 @@
 
 
 
-## How to install
+## How to Install
 
-Please refer to [Original 3DGS code base](https://github.com/graphdeco-inria/gaussian-splatting) for installation instructions. Make sure to reinstall diff-gaussian-rasterization with the following command on an available 3DGS environment as this library has been modified:
-```
-pip install submodules/diff-gaussian-rasterization
-```
+This project is built on top of the [Original 3DGS code base](https://github.com/graphdeco-inria/gaussian-splatting) and has been tested only on Ubuntu 20.04. If you encounter any issues, please refer to the [Original 3DGS code base](https://github.com/graphdeco-inria/gaussian-splatting) for installation instructions.
+
+### Installation Steps
+
+1. **Clone the Repository:**
+   ```sh
+   git clone --recursive git@github.com:ubc-vision/3dgs-mcmc.git
+   cd 3dgs-mcmc
+   ```
+2. **Set Up the Conda Environment:**
+    ```sh
+    conda create -y -n 3dgs-mcmc-env python=3.8
+    conda activate 3dgs-mcmc-env
+    ```
+3. **Install Dependencies:**
+    ```sh
+    pip install plyfile tqdm torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117
+    conda install cudatoolkit-dev=11.7 -c conda-forge
+    ```
+4. **Install Submodules:**
+    ```sh
+    CUDA_HOME=PATH/TO/CONDA/envs/3dgs-mcmc-env/pkgs/cuda-toolkit/ pip install submodules/diff-gaussian-rasterization submodules/simple-knn/
+    ```
+### Common Issues:
+1. **Access Error During Cloning:**
+If you encounter an access error when cloning the repository, ensure you have your SSH key set up correctly. Alternatively, you can clone using HTTPS.
+2. **Running diff-gaussian-rasterization Fails:**
+You may need to change the compiler options in the setup.py file to run both the original and this code. Update the setup.py with the following extra_compile_args:
+    ```sh
+    extra_compile_args={"nvcc": ["-Xcompiler", "-fno-gnu-unique", "-I" + os.path.join(os.path.dirname(os.path.abspath(__file__)), "third_party/glm/")]}
+    ```
+    Afterwards, you need to reinstall diff-gaussian-rasterization. This is mentioned in [3DGS-issue-#41](https://github.com/graphdeco-inria/gaussian-splatting/issues/41).
+
+By following these steps, you should be able to install the project and reproduce the results. If you encounter any issues, refer to the original 3DGS code base for further guidance.
 
 ## How to run
 Running code is similar to the [Original 3DGS code base](https://github.com/graphdeco-inria/gaussian-splatting) with the following differences:
@@ -30,6 +96,11 @@ Running code is similar to the [Original 3DGS code base](https://github.com/grap
 - You need to specify the noise learning rate. This is performed using --noise_lr argument. For all the experiments in the paper, we use 5e5.
 - You need to specify the initialization type. This is performed using --init_type argument. Options are random (to initialize randomly) or sfm (to initialize using a pointcloud).
 
+## How to Reproduce the Results in the Paper
+```sh
+python train.py --source_path PATH/TO/Shape --config configs/shape.json --eval
+```
+
 
 
 

arguments/__init__.py

@@ -53,7 +53,7 @@ def __init__(self, parser, sentinel=False):
         self._resolution = -1
         self._white_background = False
         self.data_device = "cuda"
-        self.eval = True
+        self.eval = False
         self.cap_max = -1
         self.init_type = "random"
         super().__init__(parser, "Loading Parameters", sentinel)

configs/bicycle.json

@@ -0,0 +1,4 @@
+{
+    "resolution": 4,
+    "cap_max": 5900000
+}

configs/bonsai.json

@@ -0,0 +1,4 @@
+{
+    "resolution": 2,
+    "cap_max": 1300000
+}

configs/counter.json

@@ -0,0 +1,4 @@
+{
+    "resolution": 2,
+    "cap_max": 1200000
+}

configs/garden.json

@@ -0,0 +1,4 @@
+{
+    "resolution": 4,
+    "cap_max": 5200000
+}

configs/kitchen.json

@@ -0,0 +1,4 @@
+{
+    "resolution": 2,
+    "cap_max": 1800000
+}

configs/room.json

@@ -0,0 +1,4 @@
+{
+    "resolution": 2,
+    "cap_max": 1500000
+}

configs/stump.json

@@ -0,0 +1,4 @@
+{
+    "resolution": 4,
+    "cap_max": 4750000
+}

environment.yml

@@ -4,13 +4,12 @@ channels:
   - conda-forge
   - defaults
 dependencies:
-  - cudatoolkit=11.6
+  - cudatoolkit=11.7
   - plyfile
-  - python=3.7.13
-  - pip=22.3.1
-  - pytorch=1.12.1
-  - torchaudio=0.12.1
-  - torchvision=0.13.1
+  - python=3.8
+  - pytorch=1.13.1
+  - torchaudio=0.13.1
+  - torchvision=0.14.1
   - tqdm
   - pip:
     - submodules/diff-gaussian-rasterization

train.py

@@ -10,6 +10,7 @@
 #
 
 import os
+import json
 import torch
 from random import randint
 from utils.loss_utils import l1_loss, ssim
@@ -203,14 +204,18 @@ def training_report(tb_writer, iteration, Ll1, loss, l1_loss, elapsed, testing_i
             tb_writer.add_scalar('total_points', scene.gaussians.get_xyz.shape[0], iteration)
         torch.cuda.empty_cache()
 
+def load_config(config_file):
+    with open(config_file, 'r') as file:
+        config = json.load(file)
+    return config
+
 if __name__ == "__main__":
     # Set up command line argument parser
     parser = ArgumentParser(description="Training script parameters")
     lp = ModelParams(parser)
     op = OptimizationParams(parser)
     pp = PipelineParams(parser)
-    parser.add_argument('--ip', type=str, default="127.0.0.1")
-    parser.add_argument('--port', type=int, default=6009)
+    parser.add_argument('--config', type=str, default=None)
     parser.add_argument('--debug_from', type=int, default=-1)
     parser.add_argument('--detect_anomaly', action='store_true', default=False)
     parser.add_argument("--test_iterations", nargs="+", type=int, default=[7_000, 30_000])
@@ -219,6 +224,14 @@ def training_report(tb_writer, iteration, Ll1, loss, l1_loss, elapsed, testing_i
     parser.add_argument("--checkpoint_iterations", nargs="+", type=int, default=[])
     parser.add_argument("--start_checkpoint", type=str, default = None)
     args = parser.parse_args(sys.argv[1:])
+
+    if args.config is not None:
+        # Load the configuration file
+        config = load_config(args.config)
+        # Set the configuration parameters on args, if they are not already set by command line arguments
+        for key, value in config.items():
+            setattr(args, key, value)
+
     args.save_iterations.append(args.iterations)
 
     print("Optimizing " + args.model_path)