naver / dust3r Goto Github PK

Hello！Thank you for open sourcing such a great project！
I want to know how to let the program output the pose, depth maps, and point clouds？

There's calling print() with passing force=True (at losses.py, L222).
However, built-in print() function can't accept this parameter:

>>> print("something", force=True)
Traceback (most recent call last):
  File "C:\Program Files\JetBrains\PyCharm 2022.3.2\plugins\python\helpers\pydev\pydevconsole.py", line 364, in runcode
    coro = func()
           ^^^^^^
  File "<input>", line 1, in <module>
TypeError: 'force' is an invalid keyword argument for print()

I believe the author implied that parameter to flush stdout stream, so there should have been flush=True:

>>> print("something", flush=True)
something

Would it be possible to use spase or dense matchers to get the scale out? Getting all the matched points and then using the known intrinsics to project them into 3d and then comparing the same points to pointmap in dust3r?

Hi, I get some f-string error, like:

dust3r/heads/__init__.py line:19
raise NotImplementedError(f"unexpected {head_type=} and {output_mode=}")

 File "<fstring>", line 1
    (head_type=)
              ^
SyntaxError: invalid syntax

Am I using the wrong approach?

When will your dataset be available for release? thanks

Fantastic work guys ! Quality and performance is impressive, though the visualizer seems to apply some transparency in mesh mode compared to point cloud.

Here is a comparaison of the 2 using a single image as source :

Also when saving the mesh there is no texture embedded and no color information on the point cloud when I import it to Blender

PS : I installed dust3r using Pinokio.

Hello,
Does dust3r reconstruct objects with metric scale accuracy? I want to tryout dust3r for a volume estimation use case and was wondering if this is possible.

Firstly, congrats to all the folks at Naver for their awesome accomplishments with Dust3r. It was very straight forward getting dust3r up and running, but I discovered i run into torch.cuda.OutOfMemoryError(s) when i try to process more than 16 images at once. I am running a rtx 3060 12GB, was wondering if anyone may know what I can do to resolve or debug an issue like this? I am running dust3r via docker-compose with PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:128. Here is the full error:

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.11 GiB (GPU 0; 11.76 GiB total capacity; 9.84 GiB already allocated; 932.69 MiB free; 10.32 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Any help or insight would be much appreciated. Again, thanks to the folks at Naver for their awesome work and releasing it!

Hi, I have a problem calling the preset_pose function of the PointCloudOptimizer class, I don't find the _get_msk_indices() function in the class or the parent class, can you please tell me the way to write this function or the code?

Would it be possible to use existing depth maps, such as from an iPhone's lidar or from a stereo camera, rather than doing monocular estimation?

Good jobs guys ! Very impressive results.

I confirm that it's working on Apple Mac ( with Apple Silicon ), I try with more than 8 images with no error.
3 images, 6 images pairs, run in 6 sec + 15 sec
8 images, 56 images pairs, run in 70 sec + 90 sec

PYTORCH_ENABLE_MPS_FALLBACK=1 python3 demo.py --weights checkpoints/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth --device mps

Did you plan to release some samples with a more bigger number of images ?

Thanks for the engineering code, I want to deploy the model to embedded devices, is this feasible, do you know those devices support this kind of model Looking forward to your reply

Testing on jpg. Example included.

Hello,

Thank you so much for the amazing work. I am wondering what convention you are using for DUST3R for the training and inference?

There seems to be an issue with the camera's intrinsics, some focal is 0.

Hi,

The performance is really amazing on the few image pairs I have tried.
However, when I moved to a bigger scenes (29 images), it crashes with CUDA OOM on 16Gb V100.
Any recommendations how can I run it?

  File "/home/old-ufo/dev/dust3r/dust3r/cloud_opt/optimizer.py", line 176, in forward
    aligned_pred_i = geotrf(pw_poses, pw_adapt * self._stacked_pred_i)
  File "/home/old-ufo/dev/dust3r/dust3r/utils/geometry.py", line 86, in geotrf
    pts = pts @ Trf[..., :-1, :] + Trf[..., -1:, :]

torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.38 GiB. GPU 0 has a total capacity of 15.77 GiB of which 775.88 MiB is free. Including non-PyTorch memory, this process has 15.01 GiB memory in use. Of the allocated memory 13.70 GiB is allocated by PyTorch, and 922.14 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

As I understand here #44 (comment) ya'll are in the process of training a metric version of dust3r, so the current version outputs scale invariant. Was the depth normalized before training? I see here

Is there any normalization being performed on the depth/pointmap to make it scale invariant?

Also I've been following what the MetricV2, have you all looked at including surface normals as supervision based on pointmap -> depthmap -> surface normal conversion such that the network can also produce surface normals?

Assuming we have 1296 images (36x36) around the object in a 360x360 setup, or perhaps slightly fewer, say 100 images.
The current algorithm generates pairs for all possible combinations, resulting in a substantial number of pairs (100*99).
Without a proper sampling policy, this can be challenging to handle.
A small recommendation would be to initiate and maintain a pair-loss table from scratch, gradually increasing the number of pairs and sampling based on the convergence of loss.

First, congrats to the developers. a powerful tool, found immediate use for my projects.

Because most of our tools are written in MATLAB/Octave, I found that the generated .glbfile is difficult to convert.

I just added a little bit code to allow the demo.py GUI to export data to a JSON/binary JSON construct (with the JMesh mesh data annotation) - which can be potentially parsed/shared in other environments (like JavaScript/Node/C++/MATLAB). I also added a drop down menu in the demo to let user choose output format.

Here are my commits

NeuroJSON@a13b18c
NeuroJSON@935ace7

to export to JSON, only one extra dependency jdata (16 kb) is needed. To export to a binary JSON format (for smaller file sizes), another small package bjdata (65 kb).

loading the data in MATLAB/Octave

>> dat=loadjd('/tmp/scene.jmsh');
>> dat
dat = 
  struct with fields:

       images: [1×10 struct]
      cameras: [1×10 struct]
       meshes: [1×10 struct]
    transform: [4×4 double]

>> dat.meshes(1)
ans = 
  struct with fields:

    MeshVertex3: [196608×3 single]
       MeshTri3: [1×1 struct]

loading the data back to Python

import jdata as jd
dat=jd.load('/tmp/scene.jmsh');
>>> dat.keys()
dict_keys(['images', 'cameras', 'meshes', 'transform'])
>>> dat['meshes'][0].keys()
dict_keys(['MeshVertex3', 'MeshTri3'])
>>> dat['meshes'][0]['MeshTri3'].keys()
dict_keys(['Data', 'Properties'])
>>> dat['meshes'][0]['MeshTri3']['Data'].shape
(189024, 3)
>>> dat['meshes'][0]['MeshTri3']['Data'].dtype
dtype('int64')

in my test, the generated scene from 10 images took 43MB size in glb, 44MB in binary json (.bmsh) and 59MB for JSON (due to base64). No noticeable difference in loading/saving speed. For both JSON/binary JSON file, changing the compressor to 'lzma' could lead to much smaller file size.

just want to share this in case others see similar needs. I am also happy to create a PR if the developers are interested in adding this feature.

(venv) D:\dust3r>python demo.py --weights checkpoints/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth
Traceback (most recent call last):
File "D:\dust3r\demo.py", line 19, in
from dust3r.inference import inference, load_model
File "D:\dust3r\dust3r\inference.py", line 10, in
from dust3r.model import AsymmetricCroCo3DStereo, inf # noqa: F401, needed when loading the model
File "D:\dust3r\dust3r\model.py", line 11, in
from .heads import head_factory
File "D:\dust3r\dust3r\heads_init_.py", line 8, in
from .dpt_head import create_dpt_head
File "D:\dust3r\dust3r\heads\dpt_head.py", line 17, in
from models.dpt_block import DPTOutputAdapter # noqa
ModuleNotFoundError: No module named 'models.dpt_block'

(venv) D:\dust3r\checkpoints>dir
驱动器 D 中的卷是 Data
卷的序列号是 3C50-8BA1

D:\dust3r\checkpoints 的目录

2024/03/04 21:16

Could you please release the code about preprocess scannet data for training and inference, we can better understand the all pipeline about this dust3r? thanks

I would like to ask if it is possible to add a mask to the images currently being promoted on the network.
If so, how should I add it?

This issue must be so stupid, but I just can't find a way to add one more picture in demo.

A very good job! I have a fixed use of the scene, 4 cameras fixed to shoot an object, but I find that each time the output camera is not in the same position, can you give some prior conditions or modify the code to fix the camera pose?

Hello
Thank you for your amazing job!
Can I export .glb/.ojb file by dust3R? (for example, with your script 'Usage' in your readme)
If it is right, how can I export?
Thank you in advance!

Great work! Could the author provide the data loader for all the other datasets and maybe a subset of the other datasets?

Thanks for your amazing job! When I use three images as input in demo.py, it will occur error as shown below:

Any suggestions?

Using the default setup, large input images were being resized to 512 x 384 (using DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth). But I wanted results with higher resolution (1024 x 768). So I followed "Extending Context Window of Large Language Models via Position Interpolation" by Meta and changed only the default image_size value of 512 to 1024 inside demo.py and multiplied the variable t inside get_cos_sin method of RoPE2D of croco/models/pos_embed.py by (512/1024). This gave pretty good results, though finetuning is most likely required for better results.

Hi, thank you for your amazing job!
Could I get the real depth from your 2D-2D matching?
If I input 2 images from a stereo camera(like self-driving), according ‘real_depth = (baseline × focal) / disparity’, is that meaning the answear of title is right?
And if I input 5 images, for some reason, I want to see all matching between each images pair(for example, img2 and img5), could you tell me how to modify you code to achieve that?(I modfied the code of "Usage" of ReadMe, however, the matchings are same for all image pair...)
Thank you!

There is a typo in Demo Learning step1.
There's no quotation mark on the path next to the pretrained
The train will not work when the code is executed.
Quotations on the route have been confirmed to work well

Thanks 😀

directly following Readme and start the demo gradio, using latest code, it shows :

(dust3r) H:\qing\AIproect\dust3r>python demo.py --weights checkpoints/DUSt3R_ViTLarge_BaseDecoder_512_dpt.pth
Traceback (most recent call last):
File "H:\qing\AIproect\dust3r\demo.py", line 9, in
import gradio
File "D:\py\anaconda3\envs\dust3r\Lib\site-packages\gradio_init_.py", line 3, in
import gradio.simple_templates
File "D:\py\anaconda3\envs\dust3r\Lib\site-packages\gradio_simple_templates_init.py", line 1, in
from .simpledropdown import SimpleDropdown
File "D:\py\anaconda3\envs\dust3r\Lib\site-packages\gradio_simple_templates\simpledropdown.py", line 6, in
from gradio.components.base import FormComponent
File "D:\py\anaconda3\envs\dust3r\Lib\site-packages\gradio\components_init_.py", line 40, in
from gradio.components.multimodal_textbox import MultimodalTextbox
File "D:\py\anaconda3\envs\dust3r\Lib\site-packages\gradio\components\multimodal_textbox.py", line 28, in
class MultimodalTextbox(FormComponent):
File "D:\py\anaconda3\envs\dust3r\Lib\site-packages\gradio\component_meta.py", line 198, in new
create_or_modify_pyi(component_class, name, events)
File "D:\py\anaconda3\envs\dust3r\Lib\site-packages\gradio\component_meta.py", line 92, in create_or_modify_pyi
source_code = source_file.read_text()
^^^^^^^^^^^^^^^^^^^^^^^
File "D:\py\anaconda3\envs\dust3r\Lib\pathlib.py", line 1059, in read_text
return f.read()
^^^^^^^^
UnicodeDecodeError: 'gbk' codec can't decode byte 0xb2 in position 1972: illegal multibyte sequence

Manually adjustable camera positions after prediction within the GUI software.

Have you considered adding this? It would be a great way to correct any cameras that the algorithm may have not predicted correctly.

Love this btw. Great work so far.

Hi guys!

Congratulations for the super nice work, it works impressively!

I just made a quick adaptation for the demo to work directly on Google Colab with minor adaptations using this original repo...

Feel free if you guys want to include a (possibly modified) version of it it in the README for people with no easy access to a GPU to quickly test it with their own images.

https://colab.research.google.com/drive/1QG8LgpK0vpnJ3JEJxELaxcc6rz-KuyQn?usp=sharing

Best regards,

Dear author, if there are no inherent constraints on the input images (such as B must always be to the left of A or even stricter constraints), what is the reason behind TransformerDecoder_1/2 and Header_1/2 being required to have different weights instead of sharing weights for later information sharing? According to my limited understanding, after 'perfect' or 'sufficient' training, Decoder_1/2 and Header_1/2 should be nearly identical. In this case, what is the significance of not sharing weights?

in shortly:
if no inherent differents, and after sufficient training, decoder_header_1/2 should be nearly identical.
consider this:
input single image I to this two-path network, output diffent point-map and camera-pose.
is it reasonable? is this we wanted?

a thought experiment:
swap input image pairs A and B, maybe we will get different output performaces: worse or better.
right?

First of all: very cool work!

I have two questions regarding reproducing pairs from the datasets for training.

Habitat

Are the scenes the pairs are generated from the same as in CroCo Habitat README?
Specifically, do the 1M pairs relate in some way to the ~1.8M pairs used in Croco?

Real datasets

For CroCoV2 you provide metadata to re-generate the crops for ARKitScenes and MegaDepth.
Specifically the CroCoV2 paper mentions,

1,070,414 pairs from ARKitScenes [8], 2,014,789 pairs from MegaDepth

Do these relate to the pairs you obtained for dust3r training in some way?

If there is some metadata similar to the one for CroCo to generate these pairs that would be greatly appreciated!

Thank you!

Thank you for your excellent work. But when I try to use cam_pose, some problems arise. Specifically, for the four pictures, the following pictures can be obtained in the output of the demo,

and you can see that there are good camera poses.
But when I apply the pose parameters directly to the point cloud locally, I get the following results.

There is an unreasonable gap between each perspective.
I want to know if the way I get the pose is wrong(poses = scene.get_im_poses())? Or is it that the point cloud results displayed on the web page do not completely correspond to the pose obtained by the model? Looking forward to your reply.

Hey Naver,

First of all great work, it is very interesting to play around with!

I'm curious, if one knows a partial set of poses and focal lengths aforehand, how should one initialize the pose-graph?

Best regards

Hi,

Thanks a lot for the amazing work. Did you use the pretrained Croco model when training Dust3r? If so, could you please point out where you load the model in the training code?

Thanks in advance!

This error come from python code from the Readme file

Hi there, congrats on the fantastic work! These are amazing results.

I'm working on 3D mapping systems for robotics, and was wondering if

Given a video, can this method help with obtaining the camera parameters, and poses for each frame?

Do you guys have any scripts already for this? I see that in the example usage you have:

    # retrieve useful values from scene:
    imgs = scene.imgs
    focals = scene.get_focals()
    poses = scene.get_im_poses()

And you can do scene.get_intrinsics() which is great, but when I run this on 12 images from the replica dataset, scene.get_intrinsics() outputs 12 different intrinsic matrices, none of which really match the original camera intrinsics of the replica dataset.

Am I doing something wrong? Should I specify the scale or resolution or something else about the images at some point? The replica images are 1200x600 (w,h) but they get resized to 512 I'm assuming.

Just wondering how I should go about getting the camera parameters for a monocular rgb video, or if that's not really possible to do super accurately yet with this method.

For extra detail, I'm using the following frames from the replica dataset

    image_filenames = [
        'frame000000.jpg', 'frame000023.jpg', 'frame000190.jpg', 'frame000502.jpg',
        'frame000606.jpg', 'frame000988.jpg', 'frame001181.jpg', 'frame001374.jpg',
        'frame001587.jpg', 'frame001786.jpg', 'frame001845.jpg', 'frame001928.jpg'
    ]
...
    images = load_images(images_path_list, size=512)
    pairs = make_pairs(images, scene_graph='complete', prefilter=None, symmetrize=True)
...
    scene = global_aligner(output, device=device, mode=GlobalAlignerMode.PointCloudOptimizer)
    loss = scene.compute_global_alignment(init="mst", niter=niter, schedule=schedule, lr=lr)
...

and the output of the scene.get_intrinsics() is as follows, I'm only showing two of the matrices here, not all 12:

print(scene.get_intrinsics())
tensor([[[250.8425,   0.0000, 256.0000],
         [  0.0000, 250.8425, 144.0000],
         [  0.0000,   0.0000,   1.0000]],
...
        [[250.7383,   0.0000, 256.0000],
         [  0.0000, 250.7383, 144.0000],
         [  0.0000,   0.0000,   1.0000]]], device='cuda:0',
       grad_fn=<CopySlices>)

compared to the ground truth camera params of the replica dataset from the camera_params.json

K_given = [
    [600.0, 0, 599.5],
    [0, 600.0, 339.5],
    [0, 0, 1]
]

here is the actual camera_params.json file in case it helps

{
    "camera": {
        "w": 1200,
        "h": 680,
        "fx": 600.0,
        "fy": 600.0,
        "cx": 599.5,
        "cy": 339.5,
        "scale": 6553.5
    }
}

Also, just curious, how would I go about running this on long videos? Or is that not possible yet?

My apologies if these are too many questions! This method is really awesome, and I'm having a lot of fun using it. Thanks again for the wonderful work!

Hi, thank you for your great work!
I'm currently working on testing this on larger datasets (5-10k images) and notice that a very large amount of (V)RAM would be required to make it work. I've already generated a pairs file to reduce the number of pairs from 50M to 3M, but this still seems to be way too large. Do you have any pointers/suggestions I could try out to make it scale better?

I'm using cloud compute with 80GB VRAM and 220GB RAM so that shouldn't be an issue btw.

Thanks for your great work! It's amazing! It can be applied to many real-world scenarios. I wonder if this is a paper submitted to CVPR 2024? The format of the paper looks like it is.

Besides, could you please tell me how many GPUs it costs to train this model? Thanks very much!

as title

Tried to demo it yesterday to a few novice users using share=True in launch.
Encountered a few problems along the way.

It seems to ignore heic files (which is often the default on mobile devices).

The problematic line where you need to add a ".heic"

You also need to add a pillow-heif depency
pip install pillow-heif

And the following lines somewhere in image.py

from pillow_heif import register_heif_opener
register_heif_opener()

Other additional problems encountered en mobile devices (but probably due to gradio) :

• The 3d viewer can't easily translate the scene (no right click + drag gesture available).
• Possibility to add one more image without re-uploading everything.
• ImagePicker instead of FilePicker

I uploaded 6 pictures

Hi and thank you for for sharing this amazing work.

Just a heads up:

I had to manually install some missing deps to make main.py run. I followed the readme and installed with conda, skipping both the optional step 3 and installing the new optional_requirements.txt

einops (conda)
tqdm (via pip)
scipy (conda)
opencv-python (pip)
trimesh (pip)

pip install "pyglet<2" # use version <2

Cheers

Hi authors, thanks for sharing the amazing work. I was wondering how you use Waymo dataset for training, as the most accurate depth is from Lidar which is sparse.