作者你好，你做的实在是太棒了！因为课程要复现论文，我也不是CV领域的，所以我的问题可能有些幼稚，我想询问一下自己的图片输入之后，输出的一张长图片各个小图片是不是不同的edge map的结果？另外输出了一张被裁剪过的图是最终的结果吗？最后我还有一个小问题，我输入了一张1200W像素的图片就提醒我显卡内存不足了，我显卡内存是4G，我主要想知道为什么一张1200W像素的照片为什么能填满4G的显存，如果您有空解答我的问题，我将不胜感激！！！

Hi Zhuo, Nice work, and it helps a lot! Thanks.

But, why does it has three different sets (e.g. train_pair_edge_set_1/2/3.lst) for Multicue dataset? Are they randomly split from the corresponding train_pair_edge.lst?

To get results in Table 7 (Multicue, boundary and edge), one should run the code separately with each of the three sets and then average their edge maps for evaluation? For their five GTs(1,2,3,4,5 ), which should be used for evaluation, or all of them?

When I use python main.py --model pidinet_tiny to generate the edge images of my data set, the average edge extraction of each image can only reach about 90ms, and when using python main.py --model pidinet_tiny, it is displayed as: Threshold for ground truth: 76.800000 on BSDS_VOC, fps: 148.358999. I would like to ask what ways can be used to improve the speed of python main.py --model pidinet_tiny edge detection, such as not generating jpg and mat files?

Hello, where can I download the checkpoints file? I want to train a table5_pidinet model, but I have encountered some difficulties. Is it convenient for email to contact?

Hello, your work is perfect, but I took a look at the CPDC program you wrote today. I don't quite understand the solution of weights_c. Did something go wrong?

python main.py --model pidinet --config carv4 --sa --dil --resume --iter-size 24 -j 0 --gpu 0 --epochs 20 --lr 0.005 --lr-type multistep --lr-steps 10-16 --wd 1e-4 --savedir /path/to/table5_pidinet --datadir path/to/BSDS500 --dataset BSDS 会出现错误 No such file or directory: 'path/to/BSDS500\PASCAL/aug_gt/0.0_0/2009_003007.png' 麻烦请教一下可以帮忙看看这是哪里出错了吗

When I test the code to generate edge maps for my own images following the readme guide, an error occurs. Running the test code by: python main.py --model pidinet_converted --config carv4 --sa --dil -j 4 --datadir ./data/test --dataset Custom --evaluate ./trained_models/table5_pidinet.pth --evaluate-converted

Error:

Namespace(ablation=False, checkinfo=False, config='carv4', datadir='./data/test', datase
 epochs=20, eta=0.3, evaluate='./trained_models/table5_pidinet.pth', evaluate_converted=
=24, lmbda=1.1, lr=0.005, lr_steps=None, lr_type='multistep', model='pidinet', only_bsds
nt_freq=10, resume=False, sa=True, save_freq=1, savedir='results/savedir', seed=16383577
0.0001, workers=4)                                                                      
{'layer0': 'cd', 'layer1': 'ad', 'layer2': 'rd', 'layer3': 'cv', 'layer4': 'cd', 'layer5
, 'layer7': 'cv', 'layer8': 'cd', 'layer9': 'ad', 'layer10': 'rd', 'layer11': 'cv', 'lay
 'ad', 'layer14': 'rd', 'layer15': 'cv'}                                                
initialization done                                                                     
conv weights: lr 0.005000, wd 0.000100  bn weights: lr 0.005000, wd 0.000010    relu wei
.000000                                                                                 
cuda is not used, the running might be slow                                             
=> loading checkpoint from './trained_models/table5_pidinet.pth'                        
=> loaded checkpoint './trained_models/table5_pidinet.pth' successfully                 
{'layer0': 'cd', 'layer1': 'ad', 'layer2': 'rd', 'layer3': 'cv', 'layer4': 'cd', 'layer5
, 'layer7': 'cv', 'layer8': 'cd', 'layer9': 'ad', 'layer10': 'rd', 'layer11': 'cv', 'lay
 'ad', 'layer14': 'rd', 'layer15': 'cv'}                                                
Traceback (most recent call last):                                                      
  File "main.py", line 418, in <module>                                                 
    main(f)                                                                             
  File "main.py", line 201, in main                                                     
    model.load_state_dict(convert_pidinet(checkpoint['state_dict'], args.config))       
  File "D:\Miniconda3\envs\paintly\lib\site-packages\torch\nn\modules\module.py", line 1
                                                                                        
    self.__class__.__name__, "\n\t".join(error_msgs)))                                  
RuntimeError: Error(s) in loading state_dict for PiDiNet:                               
        Missing key(s) in state_dict: "init_block.weight", "block1_1.conv1.weight", "blo
lock1_2.conv1.weight", ....                              
        Unexpected key(s) in state_dict: "module.init_block.weight", "module.block1_1.co
ock1_1.conv2.weight", "module.block1_2.conv1.weight"...

Seems like the loaded model does not match the defined model. @zhuoinoulu Did I miss something? Very much appreciated for your help!

Hello~ The FPS of pidinet_tiny_converted and pidinet_small_converted is higher than your results when tested on my own Geforce RTX 2080Ti, which are 203.5FPS and 162.9FPS respectively. However, the FPS of pidinet_converted is only 92.9FPS, which is lower than 96FPS. And I have another question, why the speed test code do not use the torch.cuda.synchronize() to get a more resonable result?

Hi, Your paper and code save my life really! But, there are no links or instructions when I try to get the results in Table 7 (Multicue dataset). I have also downloaded the official dataset, but it could not be directly used in this project.

I am a fresh hand in Edge Detection. Would you please share with me the code for preparing Multicue?

Thank you very much!

PS my email address, [email protected]

以下是我訓練table2的參數 但是ods, ois與論文不同 請問是該修改哪裡呢 python main.py --model pidinet --config carv4 --sa --dil --resume --iter-size 24 -j 4 --gpu 0 --epochs 14 --lr 0.005 --lr-type multistep --lr-steps 8-12 --wd 1e-1 --savedir mysave --datadir data/BSDS500/HED-BSDS --dataset BSDS --only-bsds ODS=0.761, OIS=0.775

I tried to convert the .pth file to .onnx files onnx. I followed this website onnxruntime. Finally I get a numpy array with size (5,1,1,256,256). Should I get an array with 3 dimension rather than 5 dimension? Thank you.

import onnxruntime as ort
import numpy as np
import cv2
import matplotlib.pyplot as plt

img_path = "0.jpg"
input_img = cv2.imread(img_path)

mean = np.array([0.485, 0.456, 0.406]) * 255.0
scale = 1 / 255.0
std = [0.229, 0.224, 0.225]
input_blob = cv2.dnn.blobFromImage(
    image=input_img,
    scalefactor=scale,
    size=(256, 256),  # img target size
    mean=mean,
    swapRB=True,  # BGR -> RGB
    crop=False  # center crop
)
input_blob[0] /= np.asarray(std, dtype=np.float32).reshape(3, 1, 1)

ort_sess = ort.InferenceSession("pidinet_tiny_converted_new.onnx")
outputs = ort_sess.run(None, {'input': input_blob})

outputs = np.asarray(outputs)
print(outputs.shape)

I notice that in the function train() of main.py , the model are converted to training mode through model.train(). But, batch normalization or DropOut is not adopted in the network. I wonder is it necessary to use model.train(). Can I abandon it?

thanks to your excellent works. How to convert the PASCAL VOC Context dataset to the binary graph used by GT? 请问如何把 PASCAL VOC Context dataset转换为GT使用的二值图?

Thanks for your excellent job. I am a new learner of edge detection. When I evaluate my results on NYUD dataset, I found I need very very much time(Use 32 workers)... So I would like to inquire about the timing of the evaluation in order to verify that I am doing the right thing. Thank you.

Converting your model to onnx requires loading the pytorch model (including structure and parameters), could you publish a tutorial on loading your model?