SimpleRecon: 3D Reconstruction Without 3D Convolutions

This is the reference PyTorch implementation for training and testing MVS depth estimation models using the method described in

SimpleRecon: 3D Reconstruction Without 3D Convolutions

Mohamed Sayed, John Gibson, Jamie Whatson, Victor Adrian Prisacariu, Michael Firman, and Clément Godard

Paper, ECCV 2022 (arXiv pdf), Supplemental Material, Project Page

Birdseye.Live.Reconstruction.mp4

Living.Room.Birdseye.mp4

This code is for non-commercial use; please see the license file for terms. If you do find any part of this codebase helpful, please cite our paper using the BibTex below and link this repo. Thanks!

⚙️ Setup

Assuming a fresh Anaconda distribution, you can install dependencies with:

conda env create -f simplerecon_env.yml

We ran our experiments with PyTorch 1.10, CUDA 11.3, Python 3.9.7 and Debian GNU/Linux 10.

📦 Models

Download a pretrained model into the weights/ folder.

We provide the following models:

`--config`	Model	Abs Diff↓	Sq Rel↓	delta < 1.05↑	Chamfer↓	F-Score↑
`hero_model.yaml`	Metadata + Resnet Matching	0.0885	0.0125	73.16	5.81	0.671
`dot_product_model.yaml`	Dot Product + Resnet Matching	0.0941	0.0139	70.48	6.29	0.642

hero_model is the one we use in the paper as Ours

🚀 Speed

`--config`	Model	Inference Speed (`--batch_size 1`)	Inference GPU memory	Approximate training time
`hero_model`	Hero, Metadata + Resnet	130ms / 70ms (speed optimized)	2.6GB / 5.7GB (speed optimized)	36 hours
`dot_product_model`	Dot Product + Resnet	80ms	2.6GB	36 hours

With larger batches speed increases considerably. With batch size 8 on the non-speed optimized model, the latency drops to ~40ms.

📝 TODOs:

Simple scan for folks to quickly try the code, instead of downloading the ScanNetv2 test scenes.
FPN model weights.

🏃 Running out of the box!

We've now included two scans for people to try out immediately with the code. You can download these scans from here.

Steps:

Download weights for the hero_model into the weights directory.
Download the scans and unzip them to a directory of your choosing.
Modify the value for the option dataset_path in configs/data/vdr_dense.yaml to the base path of the unzipped vdr folder.
You should be able to run it! Something like this will work:

CUDA_VISIBLE_DEVICES=0 python test.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --config_file configs/models/hero_model.yaml \
            --load_weights_from_checkpoint weights/hero_model.ckpt \
            --data_config configs/data/vdr_dense.yaml \
            --num_workers 8 \
            --batch_size 2 \
            --fast_cost_volume \
            --run_fusion \
            --depth_fuser open3d \
            --fuse_color \
            --dump_depth_visualization;

This will output meshes, quick depth viz, and socres when benchmarked against LiDAR depth under OUTPUT_PATH.

This command uses vdr_dense.yaml which will generate depths for every frame and fuse them into a mesh. In the paper we report scores with fused keyframes instead, and you can run those using vdr_default.yaml. You can also use dense_offline tuples by instead using vdr_dense_offline.yaml.

See the section below on testing and evaluation. Make sure to use the correct config flags for datasets.

💾 ScanNetv2 Dataset

Please follow the instructions here to download the dataset. This dataset is quite big (>2TB), so make sure you have enough space, especially for extracting files.

Once downloaded, use this script to export raw sensor data to images and depth files.

You should change the dataset_path config argument for ScanNetv2 data configs at configs/data/ to match where your dataset is.

The codebase expects ScanNetv2 to be in the following format:

dataset_path
    scans_test (test scans)
        scene0707
            scene0707_00_vh_clean_2.ply (gt mesh)
            sensor_data
                frame-000261.pose.txt
                frame-000261.color.jpg 
                frame-000261.color.512.png (optional, image at 512x384)
                frame-000261.color.640.png (optional, image at 640x480)
                frame-000261.depth.png (full res depth, stored scale *1000)
                frame-000261.depth.256.png (optional, depth at 256x192 also
                                            scaled)
            scene0707.txt (scan metadata and intrinsics)
        ...
    scans (val and train scans)
        scene0000_00
            (see above)
        scene0000_01
        ....

In this example scene0707.txt should contain the scan's metadata and intrinsics:

    colorHeight = 968
    colorToDepthExtrinsics = 0.999263 -0.010031 0.037048 ........
    colorWidth = 1296
    depthHeight = 480
    depthWidth = 640
    fx_color = 1170.187988
    fx_depth = 570.924255
    fy_color = 1170.187988
    fy_depth = 570.924316
    mx_color = 647.750000
    mx_depth = 319.500000
    my_color = 483.750000
    my_depth = 239.500000
    numColorFrames = 784
    numDepthFrames = 784
    numIMUmeasurements = 1632

frame-000261.pose.txt should contain pose in the form:

    -0.384739 0.271466 -0.882203 4.98152
    0.921157 0.0521417 -0.385682 1.46821
    -0.0587002 -0.961035 -0.270124 1.51837

frame-000261.color.512.png and frame-000261.color.640.png are precached resized versions of the original image to save load and compute time during training and testing. frame-000261.depth.256.png is also a precached resized version of the depth map.

All resized precached versions of depth and images are nice to have but not required. If they don't exist, the full resolution versions will be loaded, and downsampled on the fly.

🖼️🖼️🖼️ Frame Tuples

By default, we estimate a depth map for each keyframe in a scan. We use DeepVideoMVS's heuristic for keyframe separation and construct tuples to match. We use the depth maps at these keyframes for depth fusion. For each keyframe, we associate a list of source frames that will be used to build the cost volume. We also use dense tuples, where we predict a depth map for each frame in the data, and not just at specific keyframes; these are mostly used for visualization.

We generate and export a list of tuples across all scans that act as the dataset's elements. We've precomputed these lists and they are available at data_splits under each dataset's split. For ScanNet's test scans they are at data_splits/ScanNetv2/standard_split. Our core depth numbers are computed using data_splits/ScanNetv2/standard_split/test_eight_view_deepvmvs.txt.

Here's a quick taxonamy of the type of tuples for test:

default: a tuple for every keyframe following DeepVideoMVS where all source frames are in the past. Used for all depth and mesh evaluation unless stated otherwise. For ScanNet use data_splits/ScanNetv2/standard_split/test_eight_view_deepvmvs.txt.
offline: a tuple for every frame in the scan where source frames can be both in the past and future relative to the current frame. These are useful when a scene is captured offline, and you want the best accuracy possible. With online tuples, the cost volume will contain empty regions as the camera moves away and all source frames lag behind; however with offline tuples, the cost volume is full on both ends, leading to a better scale (and metric) estimate.
dense: an online tuple (like default) for every frame in the scan where all source frames are in the past. For ScanNet this would be data_splits/ScanNetv2/standard_split/test_eight_view_deepvmvs_dense.txt.
offline: an offline tuple for every keyframefor every keyframe in the scan.

For the train and validation sets, we follow the same tuple augmentation strategy as in DeepVideoMVS and use the same core generation script.

If you'd like to generate these tuples yourself, you can use the scripts at data_scripts/generate_train_tuples.py for train tuples and data_scripts/generate_test_tuples.py for test tuples. These follow the same config format as test.py and will use whatever dataset class you build to read pose informaiton.

Example for test:

# default tuples
python ./data_scripts/generate_test_tuples.py 
    --data_config configs/data/scannet_default_test.yaml
    --num_workers 16

# dense tuples
python ./data_scripts/generate_test_tuples.py 
    --data_config configs/data/scannet_dense_test.yaml
    --num_workers 16

Examples for train:

# train
python ./data_scripts/generate_train_tuples.py 
    --data_config configs/data/scannet_default_train.yaml
    --num_workers 16

# val
python ./data_scripts/generate_val_tuples.py 
    --data_config configs/data/scannet_default_val.yaml
    --num_workers 16

These scripts will first check each frame in the dataset to make sure it has an existing RGB frame, an existing depth frame (if appropriate for the dataset), and also an existing and valid pose file. It will save these valid_frames in a text file in each scan's folder, but if the directory is read only, it will ignore saving a valid_frames file and generate tuples anyway.

📊 Testing and Evaluation

You can use test.py for inferring and evaluating depth maps and fusing meshes.

All results will be stored at a base results folder (results_path) at:

opts.output_base_path/opts.name/opts.dataset/opts.frame_tuple_type/

where opts is the options class. For example, when opts.output_base_path is ./results, opts.name is HERO_MODEL, opts.dataset is scannet, and opts.frame_tuple_type is default, the output directory will be

./results/HERO_MODEL/scannet/default/

Make sure to set --opts.output_base_path to a directory suitable for you to store results.

--frame_tuple_type is the type of image tuple used for MVS. A selection should be provided in the data_config file you used.

By default test.py will attempt to compute depth scores for each frame and provide both frame averaged and scene averaged metrics. The script will save these scores (per scene and totals) under results_path/scores.

We've done our best to ensure that a torch batching bug through the matching encoder is fixed for (<10^-4) accurate testing by disabling image batching through that encoder. Run --batch_size 4 at most if in doubt, and if you're looking to get as stable as possible numbers and avoid PyTorch gremlins, use --batch_size 1 for comparison evaluation.

If you want to use this for speed, set --fast_cost_volume to True. This will enable batching through the matching encoder and will enable an einops optimized feature volume.

# Example command to just compute scores 
CUDA_VISIBLE_DEVICES=0 python test.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --config_file configs/models/hero_model.yaml \
            --load_weights_from_checkpoint weights/hero_model.ckpt \
            --data_config configs/data/scannet_default_test.yaml \
            --num_workers 8 \
            --batch_size 4;

# If you'd like to get a super fast version use:
CUDA_VISIBLE_DEVICES=0 python test.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --config_file configs/models/hero_model.yaml \
            --load_weights_from_checkpoint weights/hero_model.ckpt \
            --data_config configs/data/scannet_default_test.yaml \
            --num_workers 8 \
            --fast_cost_volume \
            --batch_size 2;

This script can also be used to perform a few different auxiliary tasks, including:

TSDF Fusion

To run TSDF fusion provide the --run_fusion flag. You have two choices for fusers

--depth_fuser ours (default) will use our fuser, whose meshes are used in most visualizations and for scores. This fuser does not support color. We've provided a custom branch of scikit-image with our custom implementation of measure.matching_cubes that allows single walled. We use single walled meshes for evaluation. If this is isn't important to you, you can set the export_single_mesh to False for call to export_mesh in test.py.
--depth_fuser open3d will use the open3d depth fuser. This fuser supports color and you can enable this by using the --fuse_color flag.

By default, depth maps will be clipped to 3m for fusion and a tsdf resolution of 0.04m³ will be used, but you can change that by changing both --max_fusion_depth and --fusion_resolution

You can optionnally ask for predicted depths used for fusion to be masked when no vaiid MVS information exists using --mask_pred_depths. This is not enabled by default.

You can also fuse the best guess depths from the cost volume before the cost volume encoder-decoder that introduces a strong image prior. You can do this by using --fusion_use_raw_lowest_cost.

Meshes will be stored under results_path/meshes/.

# Example command to fuse depths to get meshes
CUDA_VISIBLE_DEVICES=0 python test.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --config_file configs/models/hero_model.yaml \
            --load_weights_from_checkpoint weights/hero_model.ckpt \
            --data_config configs/data/scannet_default_test.yaml \
            --num_workers 8 \
            --run_fusion \
            --batch_size 8;

Cache depths

You can optionally store depths by providing the --cache_depths flag. They will be stored at results_path/depths.

# Example command to compute scores and cache depths
CUDA_VISIBLE_DEVICES=0 python test.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --config_file configs/models/hero_model.yaml \
            --load_weights_from_checkpoint weights/hero_model.ckpt \
            --data_config configs/data/scannet_default_test.yaml \
            --num_workers 8 \
            --cache_depths \
            --batch_size 8;

# Example command to fuse depths to get color meshes
CUDA_VISIBLE_DEVICES=0 python test.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --config_file configs/models/hero_model.yaml \
            --load_weights_from_checkpoint weights/hero_model.ckpt \
            --data_config configs/data/scannet_default_test.yaml \
            --num_workers 8 \
            --run_fusion \
            --depth_fuser open3d \
            --fuse_color \
            --batch_size 4;

Quick viz

There are other scripts for deeper visualizations of output depths and fusion, but for quick export of depth map visualization you can use --dump_depth_visualization. Visualizations will be stored at results_path/viz/quick_viz/.

# Example command to output quick depth visualizations
CUDA_VISIBLE_DEVICES=0 python test.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --config_file configs/models/hero_model.yaml \
            --load_weights_from_checkpoint weights/hero_model.ckpt \
            --data_config configs/data/scannet_default_test.yaml \
            --num_workers 8 \
            --dump_depth_visualization \
            --batch_size 4;

👉☁️ Point Cloud Fusion

We also allow point cloud fusion of depth maps using the fuser from 3DVNet's repo.

# Example command to fuse depths into point clouds.
CUDA_VISIBLE_DEVICES=0 python pc_fusion.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --config_file configs/models/hero_model.yaml \
            --load_weights_from_checkpoint weights/hero_model.ckpt \
            --data_config configs/data/scannet_dense_test.yaml \
            --num_workers 8 \
            --batch_size 4;

Change configs/data/scannet_dense_test.yaml to configs/data/scannet_default_test.yaml to use keyframes only if you don't want to wait too long.

📊 Mesh Metrics

We use TransformerFusion's mesh evaluation for our main results table but set the seed to a fixed value for consistency when randomly sampling meshes. We also report mesh metrics using NeuralRecon's evaluation in the supplemental material.

For point cloud evaluation, we use TransformerFusion's code but load in a point cloud in place of sampling a mesh's surface.

⏳ Training

By default models and tensorboard event files are saved to ~/tmp/tensorboard/<model_name>. This can be changed with the --log_dir flag.

We train with a batch_size of 16 with 16-bit precision on two A100s on the default ScanNetv2 split.

Example command to train with two GPUs:

CUDA_VISIBLE_DEVICES=0,1 python train.py --name HERO_MODEL \
            --log_dir logs \
            --config_file configs/models/hero_model.yaml \
            --data_config configs/data/scannet_default_train.yaml \
            --gpus 2 \
            --batch_size 16;

The code supports any number of GPUs for training. You can specify which GPUs to use with the CUDA_VISIBLE_DEVICES environment.

All our training runs were performed on two NVIDIA A100s.

Different dataset

You can train on a custom MVS dataset by writing a new dataloader class which inherits from GenericMVSDataset at datasets/generic_mvs_dataset.py. See the ScannetDataset class in datasets/scannet_dataset.py or indeed any other class in datasets for an example.

🎛️ Finetuning a pretrained model

To finetune, simple load a checkpoint (not resume!) and train from there:

CUDA_VISIBLE_DEVICES=0 python train.py --config configs/models/hero_model.yaml
                --data_config configs/data/scannet_default_train.yaml 
                --load_weights_from_checkpoint weights/hero_model.ckpt

Change the data configs to whatever dataset you want to finetune to.

🔧 Other training and testing options

See options.py for the range of other training options, such as learning rates and ablation settings, and testing options.

✨ Visualization

Other than quick depth visualization in the test.py script, there are two scripts for visualizing depth output.

The first is visualization_scripts/visualize_scene_depth_output.py. This will produce a video with color images of the reference and source frames, depth prediction, cost volume estimate, GT depth, and estimated normals from depth. The script assumes you have cached depth output using test.py and accepts the same command template format as test.py:

# Example command to get visualizations for dense frames
CUDA_VISIBLE_DEVICES=0 python ./visualization_scripts/visualize_scene_depth_output.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --data_config configs/data/scannet_dense_test.yaml \
            --num_workers 8;

where OUTPUT_PATH is the base results directory for SimpleRecon (what you used for test to begin with). You could optionally run .visualization_scripts/generate_gt_min_max_cache.py before this script to get a scene average for the min and max depth values used for colormapping; if those aren't available, the script will use 0m and 5m for colomapping min and max.

The second allows a live visualization of meshing. This script will use cached depth maps if available, otherwise it will use the model to predict them before fusion. The script will iteratively load in a depth map, fuse it, save a mesh file at this step, and render this mesh alongside a camera marker for the birdseye video, and from the point of view of the camera for the fpv video.

# Example command to get live visualizations for mesh reconstruction
CUDA_VISIBLE_DEVICES=0 python visualize_live_meshing.py --name HERO_MODEL \
            --output_base_path OUTPUT_PATH \
            --config_file configs/models/hero_model.yaml \
            --load_weights_from_checkpoint weights/hero_model.ckpt \
            --data_config configs/data/scannet_dense_test.yaml \
            --num_workers 8;

By default the script will save meshes to an intermediate location, and you can optionally load those meshes to save time when visualizing the same meshes again by passing --use_precomputed_partial_meshes. All intermediate meshes will have had to be computed on the previous run for this to work.

📝🧮👩‍💻 Notation for Transformation Matrices

This repo uses the notation "cam_T_world" to denote a transformation from world to camera points (extrinsics). The intention is to make it so that the coordinate frame names would match on either side of the variable when used in multiplication:

cam_points = cam_T_world @ world_points

world_T_cam denotes camera pose (from cam to world coords). ref_T_src denotes a transformation from a source to a reference view.

🗺️ World Coordinate System

This repo is geared towards ScanNet, so while its functionality should allow for any coordinate system (signaled via input flags), the model weights we provide assume a ScanNet coordinate system. This is important since we include ray information as part of metadata. Other datasets used with these weights should be transformed to the ScanNet system. The dataset classes we include will perform the appropriate transforms.

🙏 Acknowledgements

We thank Aljaž Božič of TransformerFusion, Jiaming Sun of Neural Recon, and Arda Düzçeker of DeepVideoMVS for quickly providing useful information to help with baselines and for making their codebases readily available, especially on short notice.

The tuple generation scripts make heavy use of a modified version of DeepVideoMVS's Keyframe buffer (thanks again Arda and co!).

The PyTorch point cloud fusion module at torch_point_cloud_fusion code is borrowed from 3DVNet's repo. Thanks Alexander Rich!

We'd also like to thank Niantic's infrastructure team for quick actions when we needed them. Thanks folks!

Mohamed is funded by a Microsoft Research PhD Scholarship (MRL 2018-085).

📜 BibTeX

If you find our work useful in your research please consider citing our paper:

@inproceedings{sayed2022simplerecon,
  title={SimpleRecon: 3D Reconstruction Without 3D Convolutions},
  author={Sayed, Mohamed and Gibson, John and Watson, Jamie and Prisacariu, Victor and Firman, Michael and Godard, Cl{\'e}ment},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2022},
}

linzhenyuyuchen/simplerecon