ai-liu / complex-yolo Goto Github PK
View Code? Open in Web Editor NEWComplex-YOLO: Real-time 3D Object Detection on Point Clouds pytorch Darknet
Complex-YOLO: Real-time 3D Object Detection on Point Clouds pytorch Darknet
How do I train on own data, I already generated data/labels, but in my case I have 3 classes. So I decided to run and changed the
region_loss = RegionLoss(num_classes=8, num_anchors=5)
to
region_loss = RegionLoss(num_classes=3, num_anchors=5)
and got an error:
RuntimeError: shape '[12, 5, 10, 16, 16]' is invalid for input of size 230400
seems like second and third element of the output shape should be equal to 75 if multiplied together. But not sure what should I change to introduce different amount of classes.
Thank you very much for the code you provided, and I've been working on point cloud detection recently, so I sincerely hope to get the model files after your training. Thank you very much.
Hi @AI-liu ,
Using the same hyperparameters from the repo, I trained it for 200 epochs on KITTI. It overfits to the train data. After training, I checked the first 100 images output using eval.py. It only gives outputs for train dataset and not for validatation dataset. Did you use same code and hyperparams for your results in the repo?
Thanks.
Thanks for you share.however,I cann't load you pretrained weights.
楼主您好,我觉得这篇论文最大的创新就是在yolo上增加了角度预测,但是看您的结果预测框并没有角度
can I compile it on windows system?
Hi, I want to convert your yolov2.weights
to tensorflow .pb
file, can you give me your .cfg
file?
Hi @AI-liu ,
Did you evaluate mAP on the KITTI? were you able to get the scores mentioned in the paper?
so i need to train the network but got confused to acquire labels , how did you get labels on Projection of LIDAR frame ?
Hello everyone, can someone please suggest me point cloud annotation tool to label my lidar data with output similar to Kitti dateset format....
Thank you.
Hi Al-liu,
thank you for your implementation!
According to the cfg file in your repository the model should have 46935494 trainable parameters. However, your weight file contains only 46931945 weights. That is a difference of 3549.
It would be great if you could explain.
Many thanks!
The Complex-YOLO paper shows oriented bounding boxes in bird view images and the authors injects a predefined height for every object. Do you finish the orientation regression and injecting the predefined height?
Thank for your source code and can we use the network directly to deal with the single-line radar point cloud data?
I am curious about how you process rotated rectangle iou in your code?
best.
Hi,
Thanks for sharing this amazing work,
but it seems the pre-trained model share link expired and was cancelled,
Is it convenient to provide a link?
Thank you very much!
Hello @AI-liu,
You have done wonderful work. The pretrained model you provided is based on DarkNet training. How about training with this repo pytorch code. Is is incomplete? I was training with your code, but couldn't get good results. Please share your results.
Thanks.
The function boxed_cam_to_velo
in file "utils.py" projects the position of ground truth 3D bounding box to velo coordinate just with calibration matrix Tr_velo_to_cam
. Does that mean the position is in camera 0 coordinate rather than rectified camera 0 coordinate? I have read some codes, such pointnet and frustum pointnets, which consider the position in rectified camera 0 coordinate.
Hello,
Thank you for sharing such a good implentation of Complex-YOLO :-)
after run python3 main.py
, I got 20's of ComplexYOLO_epochxxx
and I ran python3 eval.py
after modifying data path but Error occurred like below
Traceback (most recent call last):
File "eval.py", line 93, in <module>
target, target_num= get_target(label_file, calib['Tr_velo2cam'])
ValueError: too many values to unpack (expected 2)
any idea to resolve it?
There are 7 classes : https://github.com/AI-liu/Complex-YOLO/blob/master/utils.py#L11
But you have written code to accept 8 classes. With 7 classes the output tensor channels will be (6+1+7) x 5 = 70. In this implementation output tensor channels are 75 (https://github.com/AI-liu/Complex-YOLO/blob/master/complexYOLO.py#L74). Am I missing something or this is a mistake?
Thanks.
Could you please share a direct link to the pre-trained model weights file (for your PyTorch implementation)?
Thanks!
In https://github.com/AI-liu/Complex-YOLO#result, predicted box and predicted class is drawn.
But, I think that eval.py draws only predicted box and target box. So, I want to know how to get predicted class from all_boxes
.
Hello everyone, in the middle of training i am facing the following issue. please help if anyone knows about this problem:
File "/media/imad/Windows/3d folder/Complex-YOLOv3/utils/utils.py", line 346, in build_targets
best_ious, best_n = ious.max(0)
RuntimeError: cannot perform reduction function max on tensor with no elements because the operation does not have an identity
It's a bit difficult to download weight using pan.baidu link for the foreigner
I publicly shared my NAS link to download this weight.
It'll be expired at 25/08/2019
https://gofile.me/3TjkG/u3XtTjGTe
(link is available by 2021.06)
The original paper can predict precise 3d box, does this 2d box can generat 3d box?
How to handle rotated objects using anchors?
hi @AI-liu thanks for the wonderful code . Just had a few queries
Thanks in advance
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.