syscv / r3d3 Goto Github PK

hello! I want to learn this project, as a beginner, in order to generate training data , but my Gpu memory is not enough, 8GB only, which configuration parameters need to modify, or how much graphic memory is needed at least.
Thank you for your reply

Thank you for your work. I have encountered the following problems. It seems that the DDADDataset.py file is missing in the efm_datasets folder.

First, thank you for your excellent work.
Second, I would like to ask how to generate the ply point cloud file from the result files (.npz)? As shown in Fig. 11 and 12 in your paper for visualization results, is there a conversion script? Thank you!

Hi there, love your work! I want to ask whether we need Eigen3 to run the code, and which version should we used

Hi, thanks for opening this great work!

I wonder why the command to generate training data in step 1 is almost the same as the Evaluation process? And why the step for generating training data is so slow, and it will evaluate the metrics like the below for the nuScenes dataset?

Are there any ideas to accelerate the training data generation step?

Hello, thank you very much for your paper. I am very interested in your work. I would like to know if the reconstructed scene can be measured and if it has measurable scale information.

I have a problem when I run the code:
python evaluate.py \ --config configs/evaluation/dataset_generation/dataset_generation_ddad.yaml \ --r3d3_weights=data/models/r3d3/r3d3_finetuned.ckpt \ --r3d3_image_size 384 640 \ --r3d3_n_warmup=5 \ --r3d3_optm_window=5 \ --r3d3_corr_impl=lowmem \ --r3d3_graph_type=droid_slam \ --training_data_path=./data/datasets/DDAD

Holle, first of all, thank you for open-sourcing this work! I am interested in reproducing this work, but I encountered an issue when trying to create the virtual environment using the command (git clone --recurse-submodules https://github.com/AronDiSc/r3d3.git) as the third-party submodule (r3d3/thirdparty/eigen @ 2873916) is no longer accessible. Is there an alternative available?

Hello, thanks for your great work!

There are some outlier points that are hard to remove, when I use the '.ckpt' file you provided and visualize the result using Open3d in nuscens-0268.

May I ask how you dealt with them?

The photos below have shown the result after using different levels of uniform_down_sample and remove_radius_outlier.

And I found that these outliers appear in car because depth prediction of the overleap part of two camera is not well on a frame.

Hello, I encountered a problem when evaluating the model. It seems that the pre-trained model is different. Can you help me solve it?

Hi, thank you for releasing this great interesting work!

BTW, could you please share the trajectory estimation result on each camera, of DDAD dataset (especially CAMERA_01)?
To my understanding, their x6 camera average achieves 0.433 following your paper.

Thank you very much!

Great job!
I'd like to ask, have you tried using only spatial and temporal edges? Would that be sufficient?
Meanwhile, removing spatial-temporal edges could save a lot of computation.

Hello, thank you very much for your code. I am very interested in your work and I have tried using evaluate.py to generate nuScenes training data. But, I have noticed that its speed is a bit slow (only one GPU is in use). May I ask if there is any solution?

I tried to modify a single GPU to multiple GPUs, but failed: all GPUs are processing the same scene. May I know how to modify this code? Looking forward to your reply!

Hello, thanks for your code!
I'm blocked in the visualization step, because I'm using remote server to run this code and can't return the online visualization.
Therefore, I want to use pred result (e.g., disp, pose of every frame) for visualization via Open3D.
But it seems that that my visualization method can't align multi-frame images.

Two keyframe (interveral 5 frame) in nuscenes-0268

I'm not quite sure what step went wrong.

# Visualization Code

def processing(disp, K, T, image):
    depth = 1.0 / disp.squeeze()    # h, w
    height, width = depth.shape
    img = cv2.resize(cv2.cvtColor(image, cv2.COLOR_BGR2RGB), dsize=(
        width, height), interpolation=cv2.INTER_LINEAR)
    
    fx, fy, cx, cy = K
    intrinsics_matrix = np.array([[fx, 0, cx],
                                  [0, fy, cy],
                                  [0, 0, 1]])
    intrinsics_matrix_homo = np.zeros((3, 4), dtype=np.float32)
    intrinsics_matrix_homo[:3, :3] = intrinsics_matrix
    intrinsics_matrix_homo[2,3] = 1

    pose = pose_quaternion_to_matrix(T)

    points = []
    rgbs = []
    for v in range(height):
        for u in range(width):
            d = depth[v, u]

            if d==0: continue

            x_cam = (u - cx) / fx * d
            y_cam = (v - cy) / fy * d
            z_cam = d
            points.append([x_cam, y_cam, z_cam])
            rgbs.append(img[v, u, :])
    
    rgbs = np.array(rgbs)
    points = np.array(points)
    ones = np.ones((points.shape[0], 1))
    points = np.hstack((points, ones))  # homo

    pc_world = (pose @ points.T).T

    return points, pc_world, intrinsics_matrix_homo, pose, rgbs

cam_npy = np.load(cam_npz)
img = cv2.imread(cam_img)
disp = cam_npy['disp_up']
K = cam_npy['intrinsics']
T = cam_npy['pose']

points, pc_world, k, t, rgbs = processing(disp=disp,K=K,T=T,image=img)
pcd = o3d.geometry.PointCloud()
pcd.points = o3d.utility.Vector3dVector(pc_world[:, :3])
pcd.colors = o3d.utility.Vector3dVector(rgbs / 255.0)
o3d.visualization.draw_geometries([pcd])

Hello, I want to know whether the pre-trained model can be used to estimate the absolute depth map in a new scene, such as inputting an rgb image or a video sequence. If so, how can the scale information of multiple depth maps estimated by the pre-trained model be obtained? I want to splice multiple depth maps into a point cloud, as your video demo shows. Do you have any suggestions? I would appreciate it very much.

Hello, I'm very interested in your work! I have a question, have you tried training the network just using two frames? I've modify some of the settings and ended up with a result that doesn't look too good. May I ask if this result is normal?

Here are my modifications:

r3d3_n_warmup = 2
r3d3_optm_window = 1
r3d3_dt_intra = 1
r3d3_dt_inter = 1

And my results: