Do I need to download the wholes scannet dataset (1.2 T)?

[yyfz]

No description provided.

[chenfengxu714]

Yes, please do so as indicated by https://github.com/SamsungLabs/imvoxelnet/tree/master/data/scannet.

[yyfz]

”In this directory, extract RGB image with poses by running python extract_posed_images.py. This step is optional. Skip it if you don't plan to use multi-view RGB images. Add --max-images-per-scene -1 to disable limiting number of images per scene. ScanNet scenes contain up to 5000+ frames per each. After extraction, all the .jpg images require 2 Tb disk space. The recommended 300 images per scene require less then 100 Gb. For example multi-view 3d detector ImVoxelNet samples 50 and 100 images per training and test scene.“
Here, should I add --max-images-per-scene -1 to download all images?

[chenfengxu714]

No, you do not need to do so.

[yyfz]

No, you do not need to do so.

So we only use 300 images per scene for training, am I right?

[chenfengxu714]

yes

[yyfz]

Hi, I follow https://github.com/SamsungLabs/imvoxelnet/tree/master/data/scannet to download scannet dataset and it seems that I miss the extrinsic matrix. I get error: "extrinsic=input_dict['lidar2img'], KeyError: 'lidar2img'".

I'm sorry to bother you. I'm a beginner in 3D vision, and I'm having difficulty running your code. Hope you can give me a help :)

[Yanyirong]

Hi @yyfz
Have you successfully download the dataset?I wonder how to deal with the dataset.Do I need to download the whole 1.2T dataset first and then extract?Actually I do not have enough space now so I hope to get your help (´･ᆺ･`)

[yyfz]

Hi @yyfz Have you successfully download the dataset?I wonder how to deal with the dataset.Do I need to download the whole 1.2T dataset first and then extract?Actually I do not have enough space now so I hope to get your help (´･ᆺ･`)

Yep, this model need 300 color image and depth map per scene and there are 1700+ scenes. You can download on your server or use a portable storage device.

[chenfengxu714]

Hi, I follow https://github.com/SamsungLabs/imvoxelnet/tree/master/data/scannet to download scannet dataset and it seems that I miss the extrinsic matrix. I get error: "extrinsic=input_dict['lidar2img'], KeyError: 'lidar2img'".

I'm sorry to bother you. I'm a beginner in 3D vision, and I'm having difficulty running your code. Hope you can give me a help :)

No worry. Do you run
python tools/create_data.py scannet --root-path ./data/scannet --out-dir ./data/scannet --extra-tag scannet

[Rorsc4achhh]

Hi @chenfengxu714 , I'm just wondering that if I want to use NeRF-Det-R50's checkpoint and config file, and try to detect my own scene, how do I prepare the "ScanNetMultiViewDataset" ? Maybe it's because my own carelessness, I didn't find any info about it.
About the "posed RGB images" that NeRF-Det needs, I already got color and pose folder ready, just like what the other scannet data format look like, it's just that I seems need three pkl files that requires as mentioned in the config file you provided.

And also just wanna make sure that is this the correct way to detect a scene by using test.py?
python tools/test.py configs/nerfdet/nerfdet_res50_2x_low_res.py work_dirs/nerfdet_res50_2x_low_res/latest.pth --show --show-dir work_dirs/nerfdet_res50_2x_low_res

Sorry to bother you for these questions, I'm also quite new to this area and mmdetection3d.

[chenfengxu714]

Hi, no worry. It is a little bit tricky to customize your own dataloader under the mmdetection3D library. You should prepare camera intrinsics, extrinsics, and RGB images and you should modify ScanNetMultiViewDataset to form line 54 and line 61. So simply say, the most important items are to form input_dict with your own data.

[hollow-503]

Yes, please do so as indicated by https://github.com/SamsungLabs/imvoxelnet/tree/master/data/scannet.

Hi @chenfengxu714,
Thanks for your patient reply. I follow the tutorials to download scannet dataset.
"In this directory, extract point clouds and annotations by running python batch_load_scannet_data.py. Add the --max_num_point 50000 flag if you only use the ScanNet data for the detection task. It will downsample the scenes to less points."
Here, should I add --max_num_point 50000 to downsample the scenes? I am new in Nerf, and I don't know whether it will influence the perfomance of the detection branch.

[hollow-503]

Hi @yyfz Have you successfully download the dataset?I wonder how to deal with the dataset.Do I need to download the whole 1.2T dataset first and then extract?Actually I do not have enough space now so I hope to get your help (´･ᆺ･`)

Yep, this model need 300 color image and depth map per scene and there are 1700+ scenes. You can download on your server or use a portable storage device.

Hi @yyfz,
Here, should you add --max_num_point 50000 to downsample the scenes when preparing the Scannet dataset as mentioned in mmdet3d?