如图所示，人脸框上出现两个f分值很高的框，用阈值没法去掉，并且框也不太准，迭代了68W次，batch=96。请问这个情况是没训练好，还是训练参数没设置造成

Hi @YonghaoHe,

Thank you for sharing your code,
do you make any specific changes in the trainingsframework when training for vehicle/ pedestrian detection compared to face detection?
I am unsure on how much different object aspect ratios affect the system, as you mention in you paper that in the case of face detection we expect a near 1:1 aspect ratio. How do you modify your system to adjust for different aspect ratios?

Thx

您好，方便告诉我下Table2里的Ratio of RF and avg. face scale是怎么计算出来的么

Hoping pytorch vision bo be released

你好，在预测时候使用不同的resize_scale得到的结果很不一样。与测试后是否需要设置与训练时候相同的尺寸呢？怎么才能确保预测是resize方式与训练的时候相同。

@YonghaoHe
Sir, I have a project using Jetson Nano that requires face detection for images that are captured at night by Hikvision IP camera. I have collected several footage videos (640x360) and combined parts into one that contains faces to be detected: here . I've tried detecting faces in this video with face_recognition's find_faces_in_batches example which uses dlib's trained cnn model, but it can detect there's a face only in very few frames, faces in thousands of other frames are not detected.

I found your awesome work, and want to have a try. But my project is in a rush now, and getting around with new installation and issues is always very painful. I see that you have tested face_detection with Jetson Nano, which means you possibly already have the right environment to run. Can you please do me a favor by running my video then showing the output while I dig in finding what is going wrong with my current work? Based on that, I would know which way I should follow, I would be really really grateful for it sir.

Hi,thank you to open your code!
I use the demo in face_detection,on NVIDIA Jetson TX2, but the inference time is 70+ ms(640*480), I read your Readme.md file that inference time is 31.46ms(31.79FPS) on NVIDIA Jetson TX2，I think，I use the demo code probility works on FP32,
so I want to know that how to use FP16 in your demo on NVIDIA Jetson TX2!
thx

it was big difference between trained model with your results.

Any one knows what's the problem here? Really appreciate it !

Have you ever used labelimg to mark data? if you do this,can you share code about conveting file of xml to my train_llist.txt?
thank you !

hi Yonghao, i realize your work on pytorch , but when i run my code, the training box loss is too small i.e 5e-02 at beginning, I think the values are wrong. I just to want to know the common range for box loss and conf loss.

楼主可以分享一下训练时候的loss曲线啊

it's very slow while training the face detection,what can i do to speed up the training? the main reason for this is the pkl ，not rec for mxnet？

你好，请问下论文中的continuous face scale是怎么推出来的？为什么loss branch1的face scale是10-15？感谢。

INFO:root:Convert mxnet symbol to onnx...
INFO:root:Input shape of the model [(1, 3, 480, 640)]
INFO:root:Exported ONNX file temp.onnx saved to disk
INFO:root:Parsing onnx for trt network...
ERROR:root:Errors occur while parsing the onnx file!
ERROR:root:Error 0: Assertion failed: tensor->getDimensions().nbDims == combined->getDimensions().nbDims

我想请教一下，如果我想要训练2k的图片，但是目标只有15x15pix 大小，如何调整训练参数呢？
我尝试用v2去训练，将bbox_small_list改为[7, 12, 17]，larget_list为[12, 17, 22]，修改feature_map_size_list为计算得到的size；
训练得到的模型什么都检测不出来？是哪里存在问题吗？

`# feature map size for each scale
param_feature_map_size_list = [159, 159, 79, 79, 39, 19, 19, 19]

bbox lower bound for each scale

param_bbox_small_list = [10, 15, 20, 40, 70, 110, 250, 400]
assert len(param_bbox_small_list) == param_num_output_scales

bbox upper bound for each scale

param_bbox_large_list = [15, 20, 40, 70, 110, 250, 400, 560]
assert len(param_bbox_large_list) == param_num_output_scales

bbox gray lower bound for each scale

param_bbox_small_gray_list = [math.floor(v * 0.9) for v in param_bbox_small_list]

bbox gray upper bound for each scale

param_bbox_large_gray_list = [math.ceil(v * 1.1) for v in param_bbox_large_list]

the RF size of each scale used for normalization, here we use param_bbox_large_list for better regression

param_receptive_field_list = param_bbox_large_list

RF stride for each scale

param_receptive_field_stride = [4, 4, 8, 8, 16, 32, 32, 32]

the start location of the first RF of each scale

param_receptive_field_center_start = [3, 3, 7, 7, 15, 31, 31, 31]`

大佬您好：
　　首先，感谢开源．看完ｆａｃｅdetection里config_farm和data_iterator_farm里的内容后，关于感受野有点疑问：１，您实际训练中，８个ｂｒａｎｃｈ中每个ｂｒａｎｃｈ的ＲＦ的大小实际不是用的感受野计算公式逐层迭代得到，而是直接引用了每个尺度的上边界作为该ｂｒａｎｃｈ的ＲＦ大小，是处于ＥＲＦ的考量吗？
２，起始的感受也中心 [3, 3, 7, 7, 15, 31, 31, 31]是怎么得到的？（按照网上说的起始感受野位置计算公式center_out=center_in + ((kernel - 1）/2 +p)*不包括当前层之前的ｓｔｒｉｄｅ累积，计算的话，比这大多了．．

希望大佬能不吝赐教，谢谢！

Mxnet:1.5.1

Opencv:3.4.6

When I train pedestrian_detection model on my custom data, I got the info of Opencv error.

OpenCV(3.4.6) /io/opencv/modules/imgproc/src/resize.cpp:3721: error: (-215:Assertion failed) inv_scale_x > 0 in function 'resize'

OpenCV(3.4.6) /io/opencv/modules/imgproc/src/resize.cpp:3721: error: (-215:Assertion failed) inv_scale_x > 0 in function 'resize'

2019-10-14 16:14:30,950[INFO]: Iter[589610] -- Time elapsed: 3.2 s. Speed: 99.2 images/s.
2019-10-14 16:14:30,953[INFO]: CE_loss_score_0: --> 30.3872
2019-10-14 16:14:30,954[INFO]: SE_loss_bbox_0: --> 776.0208
2019-10-14 16:14:30,955[INFO]: CE_loss_score_1: --> 43.8484
2019-10-14 16:14:30,956[INFO]: SE_loss_bbox_1: --> 1014.4573
2019-10-14 16:14:30,958[INFO]: CE_loss_score_2: --> 127.1084
2019-10-14 16:14:30,958[INFO]: SE_loss_bbox_2: --> 502.1951
2019-10-14 16:14:30,958[INFO]: CE_loss_score_3: --> 206.1356
2019-10-14 16:14:30,958[INFO]: SE_loss_bbox_3: --> 498.5898
2019-10-14 16:14:33,823[INFO]: Iter[589620] -- Time elapsed: 2.9 s. Speed: 111.7 images/s.
2019-10-14 16:14:33,824[INFO]: CE_loss_score_0: --> 44.7673
2019-10-14 16:14:33,824[INFO]: SE_loss_bbox_0: --> 503.9911
2019-10-14 16:14:33,824[INFO]: CE_loss_score_1: --> 31.8546
2019-10-14 16:14:33,824[INFO]: SE_loss_bbox_1: --> 616.3716
2019-10-14 16:14:33,825[INFO]: CE_loss_score_2: --> 129.1859
2019-10-14 16:14:33,826[INFO]: SE_loss_bbox_2: --> 493.4957
2019-10-14 16:14:33,826[INFO]: CE_loss_score_3: --> 363.3114
2019-10-14 16:14:33,826[INFO]: SE_loss_bbox_3: --> 909.4892

Hi :
when i run face detection demo.py, got the following error:

Traceback (most recent call last):
File "demo.py", line 138, in
main()
File "demo.py", line 68, in main
NMS_threshold=0.4, NMS_flag=True, skip_scale_branch_list=[])
File "../accuracy_evaluation/predict.py", line 173, in predict
score_map = numpy.squeeze(outputs[i * 2], (0, 1))
File "/usr/local/lib/python3.5/dist-packages/numpy/core/fromnumeric.py", line 1388, in squeeze
return squeeze(axis=axis)

Hi,
I'm very interested with your work in face detection, it's really good. I would like to add a feature to the lffd network to predict both bboxes and facial landmarks. I'm taking any hints on how you would proceed/if it's something possible (data is not a problem). For now my idea is to add a facial landmark loss (regression) in each loss branch and finetune the model.

https://github.com/YonghaoHe/A-Light-and-Fast-Face-Detector-for-Edge-Devices/blob/e43d53f3d1d3ce4e105131fbca7b257049ef0dec/face_detection/symbol_farm/symbol_10_560_25L_8scales_v1.py#L292

hello,I get false true bbox during test phase, could you have some else details during traing phase？

Hi,

First of all, thank you for the project, it is super good and useful.

In the face detection repo, you say that you'll try :

much lighter backbone. MobileNet and ShuffleNet series are classic choices, maybe they are

not the best. We will explore something new and touch the limit of the speed.

Do you have any ideas about when this will be ready or how I could do it myself.

Many thanks

Thanks for sharing，can we train other class？

hello,Thanks for sharing。
Can you share the specific training process?Because I want to fine-tune my dataset

你好，首先感谢你的开源。我们在你提供的backbone上进行了修改，把所有的卷积都修改为k=3，pad=1.这样的话，在计算感受野中心的时候，全部都是0.5了，取整为0的话，是否可行。还是说backbone设计的时候应该让感受野中心分布不同

你好，
车牌检测出来需要去识别，只要一个字符在检测框外识别就会错误，所以请问一下车牌检测的AP值0.989在计算时IoU_threshold设置的是多少，这个得分是保证车牌所有字符都在检测框内的计算出来的吗？

trtModel begin
1----------------------------------------------------------------
Input filename: v2.onnx
ONNX IR version: 0.0.3
Opset version: 7
Producer name:
Producer version:
Domain:
Model version: 0
Doc string:

While parsing node number 0 [Sub -> "_minusscalar1"]:
ERROR: /home/erisuser/p4sw/sw/gpgpu/MachineLearning/DIT/release/5.0/parsers/onnxOpenSource/builtin_op_importers.cpp:241 In function combineTensorsElementwise:
[8] Assertion failed: tensor->getDimensions().nbDims == combined->getDimensions().nbDims
ERROR: failed to parse onnx file

@YonghaoHe

Thank you for your work. I would like to ask if you train your own data set. Can we count the face size in our data set and delete the layers that do not exist in our data set? For example, my data set has only 320 * 240 image sizes and certainly no loss branch 8.

It is consuming about 1504 MB Gpu memory.
The MNN version is faster.
You should remove fast and lightweight from the title of this repo.

Thanks for your great job! I meet a problem when i run the configuration_10_320_20L_5scales_v2.py to train :
from ._conv import register_converters as _register_converters 2019-08-19 17:26:46,656[INFO]: Preparing before training. 2019-08-19 17:26:46,664[INFO]: Get net symbol successfully. 2019-08-19 17:27:41,030[INFO]: Prepare the data provider for all dataiter threads ---- 2019-08-19 17:27:41,030[INFO]: Dataset statistics: 12876 positive images; 44046 negative images; 56922 images in total. 2019-08-19 17:27:41,034[INFO]: MXNet Version: 1.4.1 2019-08-19 17:27:41,035[INFO]: Training settings:----------------------------------------------------------------- 2019-08-19 17:27:41,035[INFO]: param_num_val_loops:0 2019-08-19 17:27:41,035[INFO]: param_saturation_factors:{'min_factor': 0.5, 'max_factor': 1.5} 2019-08-19 17:27:41,035[INFO]: param_log_mode:w 2019-08-19 17:27:41,035[INFO]: param_blur_factors:{'sigma': 1, 'mode': 'random'} 2019-08-19 17:27:41,035[INFO]: param_enable_vertical_flip:False 2019-08-19 17:27:41,036[INFO]: param_bbox_small_list:[10, 20, 40, 80, 160] 2019-08-19 17:27:41,036[INFO]: param_receptive_field_center_start:[3, 7, 15, 31, 63] 2019-08-19 17:27:41,036[INFO]: param_num_thread_val_dataiter:1 2019-08-19 17:27:41,036[INFO]: param_optimizer_name:sgd 2019-08-19 17:27:41,036[INFO]: param_model_save_interval:100000 2019-08-19 17:27:41,036[INFO]: param_contrast_factors:{'min_factor': 0.5, 'max_factor': 1.5} 2019-08-19 17:27:41,036[INFO]: param_lr_scheduler:<mxnet.lr_scheduler.MultiFactorScheduler object at 0x7fdc6ffdea50> 2019-08-19 17:27:41,036[INFO]: param_net_input_height:640 2019-08-19 17:27:41,037[INFO]: param_trainset_pickle_file_path:../data_provider_farm/data_folder/widerface_train_data_gt_8.pkl 2019-08-19 17:27:41,037[INFO]: param_enable_random_contrast:True 2019-08-19 17:27:41,037[INFO]: param_GPU_idx_list:[0] 2019-08-19 17:27:41,037[INFO]: param_bbox_large_gray_list:[22.0, 44.0, 88.0, 176.0, 352.0] 2019-08-19 17:27:41,037[INFO]: param_neg_image_resize_factor_interval:[0.5, 3.5] 2019-08-19 17:27:41,037[INFO]: param_display_interval:100 2019-08-19 17:27:41,037[INFO]: param_valset_pickle_file_path: 2019-08-19 17:27:41,037[INFO]: param_net_input_width:640 2019-08-19 17:27:41,038[INFO]: param_save_prefix:../saved_model/configuration_10_320_20L_5scales_v2_2019-08-19-17-26-46/train_10_320_20L_5scales_v2 2019-08-19 17:27:41,038[INFO]: param_bbox_small_gray_list:[9.0, 18.0, 36.0, 72.0, 144.0] 2019-08-19 17:27:41,038[INFO]: param_momentum:0.9 2019-08-19 17:27:41,038[INFO]: param_feature_map_size_list:[159, 79, 39, 19, 9] 2019-08-19 17:27:41,038[INFO]: param_validation_interval:10000 2019-08-19 17:27:41,038[INFO]: param_enable_horizon_flip:True 2019-08-19 17:27:41,038[INFO]: param_receptive_field_list:[20, 40, 80, 160, 320] 2019-08-19 17:27:41,038[INFO]: param_neg_image_ratio:0.1 2019-08-19 17:27:41,038[INFO]: param_enable_blur:False 2019-08-19 17:27:41,039[INFO]: param_enable_random_brightness:True 2019-08-19 17:27:41,039[INFO]: param_train_metric_update_frequency:20 2019-08-19 17:27:41,039[INFO]: param_weight_decay:0.0 2019-08-19 17:27:41,039[INFO]: param_start_index:0 2019-08-19 17:27:41,039[INFO]: param_val_batch_size:20 2019-08-19 17:27:41,039[INFO]: param_optimizer_params:{'lr_scheduler': <mxnet.lr_scheduler.MultiFactorScheduler object at 0x7fdc6ffdea50>, 'learning_rate': 0.1, 'wd': 0.0, 'begin_num_update': 0, 'momentum': 0.9} 2019-08-19 17:27:41,039[INFO]: param_brightness_factors:{'min_factor': 0.5, 'max_factor': 1.5} 2019-08-19 17:27:41,040[INFO]: param_num_image_channel:3 2019-08-19 17:27:41,040[INFO]: param_num_train_loops:2000000 2019-08-19 17:27:41,040[INFO]: param_log_file_path:../log/configuration_10_320_20L_5scales_v2_2019-08-19-17-26-46.log 2019-08-19 17:27:41,040[INFO]: param_num_output_scales:5 2019-08-19 17:27:41,040[INFO]: param_receptive_field_stride:[4, 8, 16, 32, 64] 2019-08-19 17:27:41,040[INFO]: param_pretrained_model_param_path: 2019-08-19 17:27:41,040[INFO]: param_hnm_ratio:5 2019-08-19 17:27:41,040[INFO]: param_num_thread_train_dataiter:4 2019-08-19 17:27:41,041[INFO]: param_num_output_channels:6 2019-08-19 17:27:41,041[INFO]: param_learning_rate:0.1 2019-08-19 17:27:41,041[INFO]: param_bbox_large_list:[20, 40, 80, 160, 320] 2019-08-19 17:27:41,041[INFO]: param_train_batch_size:16 2019-08-19 17:27:41,041[INFO]: param_enable_random_saturation:True 2019-08-19 17:27:41,041[INFO]: param_blur_kernel_size_list:[3] 2019-08-19 17:27:41,041[INFO]: ----------------------------------------------------------------------------------- slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method slice indices must be integers or None or have an __index__ method
There is always looping the bug: slice indices must be integers or None or have an index method.
My environment is Python 2.7.15 |Anaconda, the train data you have offered is saved by a python3 format. So i use my python3 to load the pickle file and saved as a python2 format pickle file. Did this cuase the bug?

Hi @YonghaoHe,

Very nice work and thank you for sharing your code.
Do you think this framework could be extended to general multi-class detection case?

Will the C++ version be released in the near future?

I generated onnx model file for your v2 by to_onnx.py and run it on onnxruntime.

There are 10 groups of output for an inference, which are one confidence score and bbox for 5 scales. Bbox output is exactly same with mxnet inference but not confidence score.
To verify it, I modified mxnet symbol to output convolution score before softmax operation. Then I calculate softmax myself and found the result consistant with mxnet inference.

I wonder if this problem could come from onnxruntime/mxnet softmax symbol/onnx convertion?
p.s. thank you for this great work!

大佬您好：
　　首先，感谢开源．当我运行pedestrian_detection/accuracy_evaluation/predict.py,test_images文件夹下替换成其他640*480的行人图片时，检测框不在行人上，请问是什么原因。

希望大佬能不吝赐教，谢谢！

Can I test the network through the cpu version of mxnet?

hi, author
your training code of LFFD was not provided,was it?

我不太懂这个要怎么运行，是需要自己训练模型后才能？
如果可以，希望给个详细的步骤，谢谢！

你好，我在lffd-ncnn里看到一段code

//x-left-top
float mid_value = 0;
for (int j = 0; j < fea_h * fea_w; j++) {
mid_value = RF_center_Xs_mat[j] - box_map.channel(0)[j] * constant[scale_id];
x_lt_mat[j] = mid_value < 0 ? 0 : mid_value;
}
不太理解这段code坐标值的计算原理（ RF_center_Xs_mat[j] - box_map.channel(0)[j] * constant[scale_id];），能否讲解一下呢 不胜感激

Hi @YonghaoHe you have done a great job! Any plans to port this project to Android or IOS devices? Demos are needed!

Can i download validation set in .pkl format?

How the speed gap between them?

INFO:root:Init engine from ONNX file.
INFO:root:Create TensorRT builder.
INFO:root:Create TensorRT network.
INFO:root:Create TensorRT ONNX parser.
ERROR:root:Errors occur while parsing the ONNX file!
Assertion failed: tensor->getDimensions().nbDims == combined->getDimensions().nbDims

这是哪里出了问题呢？

python configuration_30_320_20L_4scales_v1.py 训练网络出现错误，找了好久原因还是没有解决，主要有三个错误，
1是
Traceback (most recent call last):
File "/home/clb/env/lib/python3.5/site-packages/mxnet/symbol/symbol.py", line 1623, in simple_bind
ctypes.byref(exe_handle)))
File "/home/clb/env/lib/python3.5/site-packages/mxnet/base.py", line 253, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [10:27:07] src/storage/storage.cc:119: Compile with USE_CUDA=1 to enable GPU usage
Stack trace:

2是：
File "/home/clb/env/lib/python3.5/site-packages/mxnet/module/executor_group.py", line 662, in _bind_ith_exec
shared_buffer=shared_data_arrays, **input_shapes)
File "/home/clb/env/lib/python3.5/site-packages/mxnet/symbol/symbol.py", line 1629, in simple_bind
raise RuntimeError(error_msg)
RuntimeError: simple_bind error. Arguments:

3是
terminate called without an active exception
Aborted (core dumped)
能否指点一下解决方法 非常感谢

I am studying your code of LFFD and here comes some question. Why in the configuration of model v1, the param receptive field center start is written in list [3, 3, 7...]. just as the paper said, branch1 has a RF size of 55, so the center could not be at (3,3), can you give me an explanation?
Thanks a lot.