juandugit / dh3d Goto Github PK

Hi JuanDuGit,

I run the test code to compute the recall but it’s recall performance is different from the paper.

According to the DH3D paper, recall @1 and @1% are 74.16, 85.30.

However, when I ran the globaldesc_extract.py with your pertained model which is in model/global/xxx ,
I got the following results:

Avg_recall :
1 : 0.7532
2 : 0.8284
3 : 0.8624
…
Avg_one_percent_retrieved :
0.8849

I would very much appreciate it if you could give me an explanation about these results. Thank you.

Best,
Ganghee

I am curious to hear about the runtime and hardware that you used. You have stated in the paper that for a point cloud with 8192 points, one forward pass took 80ms but I did not see any hardware specifications, or information about the training time, hardware and parameters.
Another question concerns the point-cloud-based relocalization process itself. I know about implementing image-based relocalization, and will be very glad to hear how you are implementing the point-cloud-based relocalization.

Much appreciated! Thanks!

Hello,

did you also validate the point cloud retrieval for DSO? As far as I can see in the paper for DSO only the point cloud registration is validated.

Regards

According to the paper, there are two the matchability score maps (S and S') in local feature detector.

But I found out that there is no use of score map S' in the detector loss function.

And in the source code, you calculated the score but didn't use.

Is the score S' unnecessary?

In my case, I am using NVIDIA Geforce RTX 2080.
But RTX 2080 only supports CUDA 10.0 above.
tensorflow/tensorflow#22900

But the tensorflow v1.9.0, v1.10.0 that you used only supports CUDA 9.0.
https://www.tensorflow.org/install/source#tested_build_configurations

Would the replacement of GPU be the only option to train DH3D?

0%| |0/2073[00:00<?,?it/s] 2021-05-20 10:48:44.249488: E tensorflow/stream_executor/cuda/cuda_blas.cc:647] failed to run cuBLAS routine cublasSgemmBatched: CUBLAS_STATUS_EXECUTION_FAILED
2021-05-20 10:48:44.249528: E tensorflow/stream_executor/cuda/cuda_blas.cc:2505] Internal: failed BLAS call, see log for details
2021-05-20 10:48:44.887998: I tensorflow/stream_executor/stream.cc:4817] stream 0x55c2525213f0 did not memzero GPU location; source: 0x7f1da27fab50
2021-05-20 10:48:44.888031: I tensorflow/stream_executor/stream.cc:4817] stream 0x55c2525213f0 did not memzero GPU location; source: 0x7f1da27fab70
2021-05-20 10:48:44.888071: E tensorflow/stream_executor/cuda/cuda_dnn.cc:2833] failed to enqueue forward batch normalization on stream: CUDNN_STATUS_EXECUTION_FAILED

InternalError (see above for traceback): Blas xGEMMBatched launch failed : a.shape=[10,512,3], b.shape=[10,3,3], m=512, n=3, k=3, batch_size=10
[[Node: MatMul = BatchMatMul[T=DT_FLOAT, adj_x=false, adj_y=false, _device="/job:localhost/replica:0/task:0/device:GPU:0"](split, _arg_R_0_0/_17)]]

I used cuda9.0 and tensorflow1.9, ubuntu18.04，Can you take a look at it for me?

Thanks for your attention
As stated in the title, I found two model files included in the source directory. One is labelled local and the other global. I know the network is trained in three stages but if I am going to calculate local and global descriptors (also salient points) in a single forward pass, which one should I use? I.e. are both files complete models or only the global one is complete?

Hi, thank you for the great work!

I intend to modify the code and run it using the tf.Session equivalent in Tensorflow C++ API.

From the function compute_local in model.py, I see that the pointcloud input for the input dict is points, which is derived from

points = tf.concat(pcdset, 0, name='pointclouds')

in the function build_graph.

However, when I import the meta graph and list the nodes, I am unable to find any node named pointclouds.

May I know how I can access it and similarly the output node names when I run my session? Apologies if my description sounds confusing as I am quite new to this.

Is this monocular localization or only lidar based?

Thanks for the great work

I have a question regarding network structure: is it possible to change to input point number without re-training the network from scratch? Thanks