I followed the instructions in the README and trained faceboxes against WIDER-FACES and after 300 epochs (14 hrs, Titan V) I end up with 98.55% vs. the pretrained model with 98.47%.

For Pascal, I get 97.05%, which is higher than both your pretrained model and the original paper.

I also noticed the yoffset you use (20% of the box height) affects the accuracy a fair bit. Changing it to 30% of box height increased Pascal to 97.10%. I notice other variants of faceboxes use + 4, which doesn't work as well. Why is this offset in there?

Is this explained solely by a newer version of Pytorch (v1.0.1) or you did not get to 300 epochs?

parser.add_argument('--confidence_threshold', default=0.05, type=float, help='confidence_threshold')

The default condifence in pytorch and caffe is different.

Is it too low in pytorch?

If I subtract the fixed mean (104,117,123) from the image, I get a much worse result on images that have a white background. Is it better to subtract the actual mean of the image being tested?

Hi: i am trying to do an inference in my picture(not from dataset), the image preprocess (orginal size input, not resize), model load, results post process is basically follow your test.py.
But the predict point is funny, which shows following: there just one point, xmin,ymin equal xmax ymax

dets [[2.1120000e+03 1.3440000e+03 2.1120000e+03 1.3440000e+03 9.8217058e-01]]

Just curious, Any ideal why this happen?

So far, there are two thing i noticed: 1, the predict point is closer to the centre point of the box that mtcnn predict, Very Strange. facebox: [2240,960,2240,960] mtcnn: [1626,503,2737,1882]

2, the orginal predict is just one point because i trace every step, found in decode step, before boxes[:, :2] -= boxes[:, 2:] / 2, boxes was like [0.x,0.x,0,0], the last two is zero.

def decode(loc, priors, variances):
   boxes = torch.cat((
        priors[:, :2] + loc[:, :2] * variances[0] * priors[:, 2:],
        priors[:, 2:] * torch.exp(loc[:, 2:] * variances[1])), 1)
    boxes[:, :2] -= boxes[:, 2:] / 2
    boxes[:, 2:] += boxes[:, :2]
    return boxes

The extra output, detection_dimension is redundant as it can be calculated in prior_box based on width and height already.

It also fails to trace, if you attempt to use the model through JIT as it cannot follow the logic flow.

@zisianw @sfzhang15 HI

配置：pytorch0.4 + python2 因为你提提供的代码是基于python3的，所以在相应位置函数加入from future import division

利用提供的训练好的检测模型，基于测试代码，得到了2845幅图片的检测结果 然后送入官方提供的检测器． 标签文件利用的是S3FD中提供的FDDB_annotation_ellipseList_new.txt 结果形式是rectangle，并没有转ellipse 检测结果已发至2位的邮箱

最终结果在False Positives=1000时，对于DiscROC，AP只有93.9915

box_utils.py脚本里的match函数在给priors分配ground truth的时候，进行了ignore hard ground truth的步骤，也就是和某个ground truth最大的iou小于0.2时， 该ground truth 会被舍弃，如果其中有被舍弃的gt，那么ground truth的index就会改变，后面再做# ensure every gt matches with its prior of max overlap的时候是不是就不对了？这部分有点不理解

i am a beginner class of detection with deep learning so i don't know how to handle anchor box. i try to change filter numbers of loc_layers(21 to 11) in multibox in faceboxes.py but i have an error "Runtime error : the expanded size of the tensor (11584) must match the exisitng size(21824) at non singleton dimension1. Target sizes : [32 ,11584 ,4]. Tensor sizes : [32 ,21824 ,1]" i think 21824 is number of default(anchor) box and 11584 is number of prediction box so what should i do to decrease number of filters?

and i have one more qusetion

this paper has anchor densification module but i don't want to use this module fully, i want to use half of anchor densification to decrease my university project's latency so i have change prior_box.py 33th line

"dense_cx = [xself.steps[k]/self.image_size[1] for x in [j+0, j+0.25, j+0.5, j+0.75]] --> dense_cx = [xself.steps[k]/self.image_size[1] for x in [j+0, j+0.5]]"

but i have an similar error above problem so i need your advice

Height values returned by the model will always be wrong for testing. This is possibly due to the image aspect ratio not being 1:1. Regardless, the width values appear to be correct and you can fix the height values by setting them equal to the width. By removing height scores, you still get the same result in AFW and PASCAL because these tests ignore width and height (using only the centre positions). Hence, [width, height] should probably be replaced by a 'size' value, which would allow for faster training/converging.

I did a test where i changed prior_box to x,y,size and ignored the fourth value of loc returned by the model. AFW and PASCAL results are unaffected.

error when running ./make.sh, any methods ？ : ^ nms/cpu_nms.c: In function ‘__Pyx__GetException’: nms/cpu_nms.c:8961:22: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_type’ tmp_type = tstate->exc_type; ^ nms/cpu_nms.c:8962:23: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_value’ tmp_value = tstate->exc_value; ^ nms/cpu_nms.c:8963:20: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_traceback’ tmp_tb = tstate->exc_traceback; ^ nms/cpu_nms.c:8964:11: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_type’ tstate->exc_type = local_type; ^ nms/cpu_nms.c:8965:11: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_value’ tstate->exc_value = local_value; ^ nms/cpu_nms.c:8966:11: error: ‘PyThreadState {aka struct _ts}’ has no member named ‘exc_traceback’ tstate->exc_traceback = local_tb; ^ error: command 'gcc' failed with exit status 1

I ran test.py and output the txt file. But I wanted to see the face boxes, drawing rectangles to mark the faces on the original image. Here comes the problem. No matter in which datasets, there always draw the wrong part. and the rate is high. By the way, I can not use numpy.ndarray(original image) instead of tensor type.

It is PASCAL dataset.

I have reimplement FaceBoxes with c++, faceboxes.cpp. It seems FaceBoxes suitable for mobile device with some simple frontal faces, but can not get good accuracy for complicated situation.

Can i use the same code on PASCAL VOC dataset where num_classes = 20, by changing the same parameter in your code, Will it work? since you have used RPN .!

@zisianw @sfzhang15

为了表达清楚，这里就用中文了！

关于train loss中的震荡现象．

自己基于提供的代码修改参数进行了3次训练，3次的差异在于batch_size和LR．每次训练时总的iterations保持一样（120K）． 唯一的差异是自己基于单GPU进行

在默认参数下，即：batch_size=32，learning rate = 0.01，epoches = 300．训练．最终结果：PASCAL：0.9630；AFW：0.9838；FDDB：0.954/0.725； 训练的cls loss和reg loss分别为下图： trian_loss_C_FaceBoxes1.pdf trian_loss_L_FaceBoxes1.pdf

修改参数，即： batch_size=16，learning rate = 0.005，epoches = 150．训练．最终结果：PASCAL：0.9628；AFW：0.9826；DDB：0.954/0.724． 得到的cls loss和reg loss分别为下图： trian_loss_C_FaceBoxes4.pdf trian_loss_L_FaceBoxes4.pdf

修改参数，即： batch_size=8，learning rate = 0.0025，epoches = 75．训练．最终结果：PASCAL：0.9567；AFW：0.9839；FDDB：0.945/0.719． 得到的cls loss和reg loss分别为下图： trian_loss_C_FaceBoxes12.pdf trian_loss_L_FaceBoxes12.pdf

可以看到不论哪种参数配置下，loss都存在着严重的震荡现象．按理来说，这种震荡对模型是会有影响的，但模型效果却很好，这个该如何去理解！！ 谢谢了！！

@zisianw HI I ran it exactly according to the training code. but I only have one card (1080TI), and the other parameters did not change at all. Training 28h. My own training loss as follows:

2020-06-01 23:31:35,613:INFO:Epoch:1/300 || Epochiter: 1/403 || Iter: 1/120900 || L: 5.1744 C: 20.3799 || LR: 0.00100000 || Batchtime: 8.5759 s || ETA: 12 days, 0:00:31
2020-06-01 23:31:36,101:INFO:Epoch:1/300 || Epochiter: 2/403 || Iter: 2/120900 || L: 4.9724 C: 17.0202 || LR: 0.00100000 || Batchtime: 0.4392 s || ETA: 14:44:56
2020-06-01 23:31:36,637:INFO:Epoch:1/300 || Epochiter: 3/403 || Iter: 3/120900 || L: 4.7130 C: 14.2861 || LR: 0.00100000 || Batchtime: 0.4849 s || ETA: 16:16:57
2020-06-01 23:31:37,171:INFO:Epoch:1/300 || Epochiter: 4/403 || Iter: 4/120900 || L: 4.6357 C: 10.1130 || LR: 0.00100000 || Batchtime: 0.4832 s || ETA: 16:13:40
2020-06-01 23:31:37,685:INFO:Epoch:1/300 || Epochiter: 5/403 || Iter: 5/120900 || L: 4.3865 C: 8.5862 || LR: 0.00100000 || Batchtime: 0.4644 s || ETA: 15:35:48
2020-06-01 23:31:38,173:INFO:Epoch:1/300 || Epochiter: 6/403 || Iter: 6/120900 || L: 4.6417 C: 7.4876 || LR: 0.00100000 || Batchtime: 0.4380 s || ETA: 14:42:26
2020-06-01 23:31:38,742:INFO:Epoch:1/300 || Epochiter: 7/403 || Iter: 7/120900 || L: 4.5764 C: 6.7580 || LR: 0.00100000 || Batchtime: 0.5216 s || ETA: 17:30:59
2020-06-01 23:31:39,315:INFO:Epoch:1/300 || Epochiter: 8/403 || Iter: 8/120900 || L: 4.5351 C: 6.1112 || LR: 0.00100000 || Batchtime: 0.4555 s || ETA: 15:17:49
2020-06-01 23:31:41,792:INFO:Epoch:1/300 || Epochiter: 9/403 || Iter: 9/120900 || L: 4.3899 C: 5.7324 || LR: 0.00100000 || Batchtime: 2.4251 s || ETA: 3 days, 9:26:15
2020-06-01 23:31:42,196:INFO:Epoch:1/300 || Epochiter: 10/403 || Iter: 10/120900 || L: 4.2871 C: 5.4239 || LR: 0.00100000 || Batchtime: 0.3506 s || ETA: 11:46:21
2020-06-01 23:31:42,732:INFO:Epoch:1/300 || Epochiter: 11/403 || Iter: 11/120900 || L: 4.3746 C: 4.6924 || LR: 0.00100000 || Batchtime: 0.4889 s || ETA: 16:25:08
2020-06-01 23:31:43,167:INFO:Epoch:1/300 || Epochiter: 12/403 || Iter: 12/120900 || L: 4.3650 C: 4.5943 || LR: 0.00100000 || Batchtime: 0.3835 s || ETA: 12:52:42
2020-06-01 23:31:43,674:INFO:Epoch:1/300 || Epochiter: 13/403 || Iter: 13/120900 || L: 4.3817 C: 4.5847 || LR: 0.00100000 || Batchtime: 0.4606 s || ETA: 15:28:00
2020-06-01 23:31:44,178:INFO:Epoch:1/300 || Epochiter: 14/403 || Iter: 14/120900 || L: 4.5922 C: 4.5324 || LR: 0.00100000 || Batchtime: 0.4664 s || ETA: 15:39:47
2020-06-01 23:31:44,721:INFO:Epoch:1/300 || Epochiter: 15/403 || Iter: 15/120900 || L: 4.0729 C: 4.5218 || LR: 0.00100000 || Batchtime: 0.4953 s || ETA: 16:37:58
2020-06-01 23:31:45,249:INFO:Epoch:1/300 || Epochiter: 16/403 || Iter: 16/120900 || L: 4.7230 C: 4.3706 || LR: 0.00100000 || Batchtime: 0.4744 s || ETA: 15:55:52
2020-06-01 23:31:48,890:INFO:Epoch:1/300 || Epochiter: 17/403 || Iter: 17/120900 || L: 4.3539 C: 4.2943 || LR: 0.00100000 || Batchtime: 3.5956 s || ETA: 5 days, 0:44:07
2020-06-01 23:31:49,361:INFO:Epoch:1/300 || Epochiter: 18/403 || Iter: 18/120900 || L: 4.0115 C: 3.8395 || LR: 0.00100000 || Batchtime: 0.4222 s || ETA: 14:10:41
2020-06-01 23:31:49,843:INFO:Epoch:1/300 || Epochiter: 19/403 || Iter: 19/120900 || L: 4.1465 C: 3.9062 || LR: 0.00100000 || Batchtime: 0.4302 s || ETA: 14:26:40
2020-06-01 23:31:50,383:INFO:Epoch:1/300 || Epochiter: 20/403 || Iter: 20/120900 || L: 4.2397 C: 3.8839 || LR: 0.00100000 || Batchtime: 0.4904 s || ETA: 16:27:56
2020-06-01 23:31:50,907:INFO:Epoch:1/300 || Epochiter: 21/403 || Iter: 21/120900 || L: 4.1348 C: 3.8818 || LR: 0.00100000 || Batchtime: 0.4766 s || ETA: 16:00:12
2020-06-01 23:31:51,456:INFO:Epoch:1/300 || Epochiter: 22/403 || Iter: 22/120900 || L: 4.2044 C: 3.6742 || LR: 0.00100000 || Batchtime: 0.5001 s || ETA: 16:47:37
2020-06-01 23:31:51,966:INFO:Epoch:1/300 || Epochiter: 23/403 || Iter: 23/120900 || L: 4.3516 C: 3.8494 || LR: 0.00100000 || Batchtime: 0.4635 s || ETA: 15:33:50
....................................................
2020-06-03 03:10:04,192:INFO:Epoch:300/300 || Epochiter: 391/403 || Iter: 120888/120900 || L: 1.0186 C: 1.5184 || LR: 0.00001000 || Batchtime: 0.5481 s || ETA: 0:00:07
2020-06-03 03:10:04,650:INFO:Epoch:300/300 || Epochiter: 392/403 || Iter: 120889/120900 || L: 1.2653 C: 1.6692 || LR: 0.00001000 || Batchtime: 0.4121 s || ETA: 0:00:04
2020-06-03 03:10:05,650:INFO:Epoch:300/300 || Epochiter: 393/403 || Iter: 120890/120900 || L: 1.3480 C: 1.7201 || LR: 0.00001000 || Batchtime: 0.9479 s || ETA: 0:00:10
2020-06-03 03:10:06,072:INFO:Epoch:300/300 || Epochiter: 394/403 || Iter: 120891/120900 || L: 1.3920 C: 1.7500 || LR: 0.00001000 || Batchtime: 0.3724 s || ETA: 0:00:03
2020-06-03 03:10:06,473:INFO:Epoch:300/300 || Epochiter: 395/403 || Iter: 120892/120900 || L: 1.2428 C: 1.8737 || LR: 0.00001000 || Batchtime: 0.3517 s || ETA: 0:00:03
2020-06-03 03:10:06,862:INFO:Epoch:300/300 || Epochiter: 396/403 || Iter: 120893/120900 || L: 0.9238 C: 1.3636 || LR: 0.00001000 || Batchtime: 0.3386 s || ETA: 0:00:02
2020-06-03 03:10:07,221:INFO:Epoch:300/300 || Epochiter: 397/403 || Iter: 120894/120900 || L: 1.4052 C: 1.8867 || LR: 0.00001000 || Batchtime: 0.3073 s || ETA: 0:00:02
2020-06-03 03:10:07,538:INFO:Epoch:300/300 || Epochiter: 398/403 || Iter: 120895/120900 || L: 1.7991 C: 2.2399 || LR: 0.00001000 || Batchtime: 0.2653 s || ETA: 0:00:01
2020-06-03 03:10:07,841:INFO:Epoch:300/300 || Epochiter: 399/403 || Iter: 120896/120900 || L: 1.3118 C: 1.8016 || LR: 0.00001000 || Batchtime: 0.2534 s || ETA: 0:00:01
2020-06-03 03:10:08,106:INFO:Epoch:300/300 || Epochiter: 400/403 || Iter: 120897/120900 || L: 1.4301 C: 2.0517 || LR: 0.00001000 || Batchtime: 0.2131 s || ETA: 0:00:00
2020-06-03 03:10:08,371:INFO:Epoch:300/300 || Epochiter: 401/403 || Iter: 120898/120900 || L: 1.5059 C: 2.0308 || LR: 0.00001000 || Batchtime: 0.2136 s || ETA: 0:00:00
2020-06-03 03:10:08,637:INFO:Epoch:300/300 || Epochiter: 402/403 || Iter: 120899/120900 || L: 0.8883 C: 1.4389 || LR: 0.00001000 || Batchtime: 0.2136 s || ETA: 0:00:00
2020-06-03 03:10:08,743:INFO:Epoch:300/300 || Epochiter: 403/403 || Iter: 120900/120900 || L: 0.9817 C: 1.6208 || LR: 0.00001000 || Batchtime: 0.0920 s || ETA: 0:00:00

It is not known whether the above loss changes are normal. Thank you!