Comments (6)
additionally, the centerpoint baseline in openpcdet is 0.5 nds lower than ours, so now there are about 0.4 nds difference.
do you use similar voxelization as ours? https://github.com/tianweiy/CenterPoint/blob/db36c497a71014961c1ec17042a7524a79d4e792/det3d/models/readers/dynamic_voxel_encoder.py#L19
My reproduction results of CenterPoint with OpenPCDet (mAP: 58.81, NDS: 66.32) are indeed slightly lower than yours.
I use the same voxelization method as yours (dynamic voxelization and pad the points in the same way). But I pad the points immediately after loading like this and it can be voxelized properly. Then I set the first 3 channels of real and virtual points to zero and do MeanVFE and scaling.
Lidar points | x | y | z | i | t | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1
Real points | x | y | z | 0 | 0 | x | y | z | c | c | c | c | c | c | c | c | c | c | s | t | 1 | 0
Virtual points | x | y | z | 0 | 0 | x | y | z | c | c | c | c | c | c | c | c | c | c | s | t | 0 | 0
I just find that my bin files in gt_database are stored as:
Lidar points | x | y | z | i | t | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 1
Real points | 0 | 0 | 0 | 0 | 0 | x | y | z | c | c | c | c | c | c | c | c | c | c | s | t | 1 | 0
Virtual points | 0 | 0 | 0 | 0 | 0 | x | y | z | c | c | c | c | c | c | c | c | c | c | s | t | 0 | 0
When they were loaded in GT_Aug, the first 3 channels of real and virtual points were still zero and were not assigned to the correct voxels. I'll fix this and see the results.
from mvp.
i didn't do any TTA for the 69.9 result
from mvp.
additionally, the centerpoint baseline in openpcdet is 0.5 nds lower than ours, so now there are about 0.4 nds difference.
do you use similar voxelization as ours? https://github.com/tianweiy/CenterPoint/blob/db36c497a71014961c1ec17042a7524a79d4e792/det3d/models/readers/dynamic_voxel_encoder.py#L19
from mvp.
@tianweiy Could you tell me your training time of MVP? I used 4 RTX 2080. CenterPoint takes about 31 hours but MVP takes about 10 days. It seems a little bit too long.
from mvp.
@tianweiy Could you tell me your training time of MVP? I used 4 RTX 2080. CenterPoint takes about 31 hours but MVP takes about 10 days. It seems a little bit too long.
Something wrong with I/O. Fix it. Now it takes 3d to train MVP with 4 RTX 3090.
from mvp.
Thanks for updating. It also takes about 2~3 days on 4 V100. I think it depends heavily on IO, CPU speed as the inference time is actually similar to vanilla centerpoint (maybe 20% slower)
from mvp.
Related Issues (20)
- Question for global_scaling_v2 HOT 2
- How long do you need to train MVP? HOT 2
- question about MaskFormer pretrained model. HOT 6
- virtual points and points features alignment HOT 9
- Generating results on nuscenes test set HOT 2
- MaskFormer pretrained on coco-panoptic HOT 1
- Where is the difference between creating data with or without virtual point? HOT 1
- Missing sweeps virtual points HOT 2
- Doubts about virtual_lidar_points HOT 1
- semantic information different between PointPainting and MVP HOT 3
- Can it be applied to the waymo dataset? HOT 3
- Sorry
- How to implement CenterPoint + Ours(w/o virtual) HOT 3
- Generage `data/nuScenes/infos_train_10sweeps_withvelo_filter_painted_True.pkl` file HOT 3
- what is the shape of the virtual points dimension? HOT 6
- generate the points HOT 1
- Is there an online implementation of MVP? HOT 1
- Visualization HOT 2
- generate the virtual points HOT 1
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from mvp.