Scaled-YOLOv4: Scaling Cross Stage Partial Network #7087

AlexeyAB · 2020-12-07T13:51:54Z

Scaled-YOLOv4: Scaling Cross Stage Partial Network - The best neural network for object detection (Top1 accuracy on MS COCO dataset)

Scaled YOLO v4 is the most accurate neural network (55.8% AP Microsoft COCO) among neural network published. In addition, it is the best in terms of the ratio of speed to accuracy in the entire range of accuracy and speed from 15 FPS to 1774 FPS. We show that YOLO and Cross-Stage-Partial (CSP) Network approaches are the best in terms of both absolute accuracy and accuracy-to-speed ratio.

Video: https://youtu.be/YDFf-TqJOFE
Paper (CVPR 2021): https://openaccess.thecvf.com/content/CVPR2021/html/Wang_Scaled-YOLOv4_Scaling_Cross_Stage_Partial_Network_CVPR_2021_paper.html
Medium: https://alexeyab84.medium.com/scaled-yolo-v4-is-the-best-neural-network-for-object-detection-on-ms-coco-dataset-39dfa22fa982?source=friends_link&sk=c8553bfed861b1a7932f739d26f487c8
Pytorch: YOLOv4-CSP, YOLOv4-P5, YOLOv4-P6, YOLOv4-P7 (use to reproduce results): https://github.com/WongKinYiu/ScaledYOLOv4
Darknet: YOLOv4-tiny, YOLOv4-CSP, YOLOv4x-MISH: https://github.com/AlexeyAB/darknet

Models:

YOLOv4-CSP (the same weights for 512x512 or 640x640)
- cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-csp.cfg
- weights: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-csp.weights
- pre-trained weights for training: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-csp.conv.142
YOLOv4x-MISH (the same weights for 512x512 or 640x640)
- cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4x-mish.cfg
- weights: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4x-mish.weights
- pre-trained weights for training: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4x-mish.conv.166
YOLOv4-P5 (the same weights for 896x896)
- cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-p5.cfg
- weights: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-p5.weights
- pre-trained weights for training: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-p5.conv.232
YOLOv4-P6 (the same weights for 1280x1280)
- cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-p6.cfg
- weights: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-p6.weights
- pre-trained weights for training: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-p6.conv.289
YOLOv4-CSPx-P7 (1536x1536)
- cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/cspx-p7-mish.cfg
- pre-trained weights for training: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/cspx-p7-mish_hp.344.conv
YOLOv4-Tiny (416x416)
- cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-tiny.cfg
- weights: hhttps://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights
- pre-trained weights for training: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.conv.29

For Training (yolov4-csp.cfg, yolov4x-mish.cfg, yolov4-p5.cfg, yolov4-p6.cfg) - change these lines before each of 3 for p5 (of 4 for p6) [yolo]-layers:

darknet/cfg/yolov4-p5.cfg

Lines 1810 to 1811 in 9a86fce

    
           filters=340 
        
           activation=logistic

filters=<(5 + num_classes) x 4>
activation=logistic - for training and detection by using Darknet: https://github.com/AlexeyAB/darknet
activation=linear - for training and detection by using Pytorch Scaled-YOLOv4 (CSP-branch): https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-csp

For training use pre-trained weights:

for yolov4-p5.cfg https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-p5.conv.232
for yolov4-p6.cfg https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-p6.conv.289

Currently Pytorch is more suitable for training on multiple-GPUs.

The text was updated successfully, but these errors were encountered:

MKiremitci · 2020-12-07T18:30:08Z

@AlexeyAB Great job sir. When I have a time, I will read the paper. Can we train a custom model with Scaled-YOLOv4?

AlexeyAB · 2020-12-07T20:57:26Z

Can we train a custom model with Scaled-YOLOv4?

Yes, sure, you can. Just use the latest version of Darknet.

MKiremitci · 2020-12-07T21:27:18Z

Thanks :) I will try it

marvision-ai · 2020-12-07T23:42:46Z

@AlexeyAB Very interesting! Great progress. Just to confirm:

Did you also update the yolov4-tiny model to be scaled-yolov4-tiny? Or is it just the regular one?
-> If yes, is there any 3layer version of it?
Does opencv support the new scaled networks?

tdurand · 2020-12-08T19:19:37Z

@AlexeyAB congrats ! same Q as @marvision-ai , did the yolov4-tiny model weights changed or is it the same ?

philipp-schmidt · 2020-12-09T01:16:19Z

Impressive work, congrats Aleksey @AlexeyAB, to you and the authors.
Just a thought of mine after reading the paper: while yolov4-tiny has an insane performance, yolov4-tiny-3l seems to be in a very nice "sweet spot", with a significant boost in accuracy over tiny (actually just a few percent shy of some of the original yolov3 variants if I can see that correctly, which is incredible if you think about how that's where it all started) for a still very impressive performance.
Is there any chance you could share pretrained weights for yolov4-tiny-3l with us so we can play around with that? I think this variant of the network really deserves more credit.

marvision-ai · 2020-12-09T19:01:56Z

@YashasSamaga Hello, does opencv support the new networks?

marvision-ai · 2020-12-14T19:37:49Z

@AlexeyAB any comments please?
@tdurand I believe there isn't any change since the cfg hasn't changed.

AlexeyAB · 2020-12-15T02:42:02Z

I have added an explanation of the necessary fixes, so we are waiting for the fix:

for OpenCV: opencv dnn module suport Scaled yolov4 model? opencv/opencv#18975 (comment)
for tkDNN: Accuracy and speed of yolov4x-mish #6987 (comment)

AlexeyAB · 2020-12-15T04:11:35Z

did the yolov4-tiny model weights changed or is it the same ?

Yes. I uploaded new weights files yolov4-tiny.weights and yolov4-tiny.conv.29 .
In general cfg-file is the same, just it was trained 2 000 000 iterations instead of 500 000 iterations.

marvision-ai · 2020-12-15T11:25:34Z

did the yolov4-tiny model weights changed or is it the same ?

Yes. I uploaded new weights files yolov4-tiny.weights and yolov4-tiny.conv.29 .
In general cfg-file is the same, just it was trained 2 000 000 iterations instead of 500 000 iterations.

I assume I can use this model with tiny-3l-spp cfg? SPP is still useful for this one correct?

philipp-schmidt · 2020-12-15T12:08:18Z

@AlexeyAB could we have weights and updated cfg for yolov4-tiny-3l.cfg as well?

Paper reports 28.7% AP (+6% to tiny) with same FPS on TX2.

AlexeyAB · 2020-12-15T20:32:16Z

I assume I can use this model with tiny-3l-spp cfg? SPP is still useful for this one correct?

You can use this weight to train your custom tiny-3l-spp: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.conv.29

marvision-ai · 2020-12-15T22:12:37Z

I assume I can use this model with tiny-3l-spp cfg? SPP is still useful for this one correct?

You can use this weight to train your custom tiny-3l-spp: https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.conv.29

@AlexeyAB Thank you.
I see there is new activations along with new features in the yolo layers. Could we have a new tiny-3l-spp cfg to train with the new weights? Thank you!

scianand · 2020-12-19T19:25:12Z

Which cfg file I have to use for Scaled-YOLOv4?

bulatnv · 2020-12-22T10:42:43Z

cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-tiny.cfg

Hello, AlexeyAB.
There seems to be an error in the mask for yolov4-tiny.

darknet/cfg/yolov4-tiny.cfg

Line 278 in c5b8bc7

mask = 1,2,3

Shouldn't it be mask=0,1,2?

Hyvenos · 2020-12-23T10:31:23Z

First of all, thanks you for you great work.
I just tried the models you provided (yolov4x-mish and yolov4-csp) to compare with the original yoloV4 on one image.
I used the latest darknet update of this repo to run my tests.

For yolov4-csp and yolov4x-mish I used the cfg and weights files you gave on the first post. For the original v4 weights and original v4 config: I used the ones you provided like two months ago.
Here are the (visual) results:

Original V4 512
yolov4x-mish 512
yolov4x-mish 640
yolov4-csp 512

Two things jumped out at me:

For csp and mish bounding boxes are more accurate, which was attented
csp and mish missed a lot of detections compared to the original v4 and it seems to be related to the object size.
AlexAB, I found a post of you saying that:
・ YOLOv4x-mish - should be trained longer. Also if you use 416x416 or 512x512, then try to use anchors from yolov4.cfg.
Which implies that the anchors are not the same between original v4 and scaled-v4.
However looking at the cfg file: I exactly have the sames.
I wonder if the small objects not detected aren't related to an anchor issue?

A last thing in which I'm less interested is that the original v4 detect the whole crowd as a person but mish and csp did not (which is suitable for my point of view)

AlexeyAB · 2020-12-24T03:46:18Z

@Hyvenos Just try to use lower confidence threshold for csp and x-mish models, -thresh 0.15

Hyvenos · 2020-12-24T08:06:14Z

Hi tried it, without improvements unfortunatelly.
The images I linked above where processed with a confidence threshold of 0.25.
I tried again with threshold 15, 10 and 5, with the same config files and weights I used in my previous post, the only thing is that I didn't test is mish in 640x640, I just test it in 512x512

Here are the results
Confidence threshold 15
Original v4
yolov'x-mish
yolov4-csp
Confidence threshold 10
Original v4
yolov4x-mish
yolov4-csp
Confidence threshold 5
Original v4
yolov4x-mish
yolov4-csp

Even when lowering the threshold, csp and mish struggle to detect the tinyest objects, and when doing they output the wrong category.
Here are the anchors:
For csp and mish: anchors = 12, 16, 19, 36, 40, 28, 36, 75, 76, 55, 72, 146, 142, 110, 192, 243, 459, 401
Can you confirm these are suitable in most of the cases?

AlexeyAB · 2020-12-24T16:42:42Z

@Hyvenos
In general, compared to yolov4 with the same resolution, the model yolov4-csp :

increases AP (bbox coordinates) ~ +4%
increases AP50 (number of detected objects) ~ +0.4%
- increases AP50 for large objects, but decreases AP50 for small objects

So:

or use yolov4-csp or yolov4x-mish with much higher resolution than yolov4
or use yolov4

If you want to detect small objects using new models csp/x-mish, then its better to fine-tune yolov4x-mish.cfg/weights on COCO or your dataset by using Darknet with these modifications for each [yolo] layer in yolov4x-mish.cfg

objectness_smooth=0
scale_x_y = 1.05
remove iou_loss=ciou and iou_normalizer=0.07
as a last resort: remove new_coords=1 (and change activation=logistic to activation=linear in [convolutional] layers before [yolo] layers)

Also if you want to detect objects in crowds, then train with iou_thresh=1.0 instead of iou_thresh=0.2

GeekAlexis · 2021-01-04T00:13:33Z

@AlexeyAB Can you elaborate a bit more on why it is better to set iou_thresh=1.0 to detect objects in crowds? Thanks in advance.

AlexeyAB · 2021-01-04T00:56:51Z

@GeekAlexis

iou_thresh=1.0 - it will use only 1 the best anchor with the highest IoU for each ground truth. So other anchors can be used for other ground truth in the same cell. So there can be 3 objects with differrent shapes in 1 cell.
iou_thresh=0.2 - solves blinking issue by matching all anchors with IoU>0.2 to 1 ground truth. It increases rewritten_boxes (value that you can see during training), so if there are 3 ground truth in 1 cell, then they will share the same 3 anchors (conflict) - all 3 anchors will be trained for 1 bbox (average between 3 ground truths).

stephanecharette · 2021-01-04T01:05:09Z

@AlexeyAB This is the kind of thing that should be documented in the wiki instead of an issue. This way people looking up details on parms can find it much easier than browsing through all the issues.

GeekAlexis · 2021-01-04T01:10:51Z

@AlexeyAB Thanks for the clarification. To train for both small and large objects, is it still recommended to use the aforementioned modifications with yolov4-csp?

AlexeyAB · 2021-01-05T15:31:11Z

yolov4-tiny don't use iou_thresh. There is very low capacity in the yolov4-tiny to train 3 instead of 1 anchor per 1 objects.

manoj-8246 · 2021-01-06T14:34:38Z

@AlexeyAB I am trying to test the scaled model on a video with below command-
./darknet.exe detector demo cfg/coco.data cfg/yolov4-csp.cfg yolov4-csp.weights sample.mp4 -out_filename output_scaled.mp4

But I am getting following message and demo screen is hanged-
OpenCV: FFMPEG: tag 0x58564944/'DIVX' is not supported with codec id 12 and format 'mp4 / MP4 (MPEG-4 Part 14)'

The same command with Yolov4 cfg and weights files works without any issue. Is there any change required for scaled version of cfg and weight files?

AlexeyAB · 2021-01-06T23:37:11Z

@manoj-8246 This is not related to YOLO. This is related to OpenCV and your video file, better to ask it there https://answers.opencv.org/questions/

haileykwak · 2021-01-20T06:04:52Z

@AlexeyAB

What is the difference between yolov4-csp and yolov4x-mish ?
and I trained darknet "yolov4-csp-416" on a custom dataset. But, Loss is very high. I checked dataset files using -show_imgs and cfg file. mAP is ok. what's the problem about loss?

RasmusToivanen · 2021-01-22T08:25:47Z

Thanks again @AlexeyAB for your great work.

If I understand the graph correctly there is no yolov4-tiny versions in the graph.
Could those be added to the graph for easier comparison against other versions.

I would also like to see "previous" yolov4-tiny models and their variants compared to scaled-yolov4 variants.
I got like 13 fps on my latest try with the "original" yolov4-tiny models on jetson nano and hopefully this weekend can get the scaled-versions running with the new model and hopefully can get the 39fps or close as the repo claims.

AlexeyAB · 2021-01-22T15:34:27Z

@RasmusToivanen There is the same yolov4-tiny.cfg model in the Darknet and Scaled version https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-tiny
Just there is available new pre-trained weights with AP=22%, AP50 = 42% (green dot on the graph) https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-tiny.weights

Small model YOLOv4-tiny isn't suitable for high-end GPU Tesla V100. YOLOv4-tiny is for Jetson / Mobile platforms.
Scaled versions of YOLOv4-CSP, P5-P7 have only +1% higher AP50 (detect the same number of objects), but has +4% higher AP (detect more accurate box coordinates).

There is comparison fo AP50 for: YOLOv4-CSP vs YOLOv4 vs YOLOv4-tiny

sharoseali · 2021-02-02T22:22:21Z

@Hyvenos
In general, compared to yolov4 with the same resolution, the model yolov4-csp :

increases AP (bbox coordinates) ~ +4%

increases AP50 (number of detected objects) ~ +0.4%

increases AP50 for large objects, but decreases AP50 for small objects

So:

or use yolov4-csp or yolov4x-mish with much higher resolution than yolov4

or use yolov4

If you want to detect small objects using new models csp/x-mish, then its better to fine-tune yolov4x-mish.cfg/weights on COCO or your dataset by using Darknet with these modifications for each [yolo] layer in yolov4x-mish.cfg

objectness_smooth=0

scale_x_y = 1.05

remove iou_loss=ciou and iou_normalizer=0.07

as a last resort: remove new_coords=1 (and change activation=logistic to activation=linear in [convolutional] layers before [yolo] layers)

Also if you want to detect objects in crowset stride=4 instead of https://github.com/AlexeyAB/darknet/blob/6f718c257815a984253346bba8fb7aa756c55090/cfg/yolov4.cfg#L989ds, then train with iou_thresh=1.0 instead of iou_thresh=0.2

HI @AlexeyAB, For improving the detection of small objects on yolov4, you mentioned 3 changes for normal Yolov4 cfg file on the readme section i-e:
for training for small objects (smaller than 16x16 after the image is resized to 416x416) - set layers = 23 instead of

darknet/cfg/yolov4.cfg

Line 895 in 6f718c2

layers = 54

set stride=4 instead of

darknet/cfg/yolov4.cfg

Line 892 in 6f718c2

stride=2

set stride=4 instead of

darknet/cfg/yolov4.cfg

Line 989 in 6f718c2

stride=2

I tried it in past for yolov4.cfg, it results in some better outcomes, but when I did some same changes in yolov4-CSP.cfg, it's giving an error while loading network at training time, seems I did a mistake .... input shape mismatch at some layer, Can you please suggest, if we want to make same changes for Scaled yolov4 (CSP) then what exactly changes I have to do? Thanks in advance.

xjinai · 2021-02-17T22:24:39Z

@AlexeyAB: Thank you for adding the support of Scaled-YOLOv4 in this repo. Great job!

I've read the Scaled-YOLOv4 paper. I'm interested in trying YOLOv4-P5, YOLOv4-P6 and YOLOv4-P7 on my custom dataset. By reading your notes above, there are models such as YOLOv4-CSP, YOLOv4x-MISH, and YOLOv4-CSPx-P7 (1536x1536). I have a couple of questions that I hope you can help:

Do you also have models for YOLOv4-P5 (896x896) and YOLOv4-P6 (1280x1280)? If not, is it possible to add those?
I am assuming that YOLOv4-CSPx-P7 (1536x1536) in your repo is the same as the YOLOv4-P7 in the paper. Can you please confirm? Can I train the YOLOv4-CSPx-P7 model using my custom dataset?
What are the differences between YOLOv4-CSP and YOLOv4x-MISH in your repo? Which one implements the "YOLOv4-CSP" in the paper?

I look forward to your reply. Thank you in advance for any information you may provide.

SpongeBab · 2021-03-22T02:15:14Z

@AlexeyAB: hi,
i use the Scaled-yolov4 to train my dataset,and i get the pt weights
how can i use it on darknet?

akashAD98 · 2021-04-21T11:03:54Z

@Hyvenos
In general, compared to yolov4 with the same resolution, the model yolov4-csp :

increases AP (bbox coordinates) ~ +4%

increases AP50 (number of detected objects) ~ +0.4%

increases AP50 for large objects, but decreases AP50 for small objects

So:

or use yolov4-csp or yolov4x-mish with much higher resolution than yolov4

or use yolov4

If you want to detect small objects using new models csp/x-mish, then its better to fine-tune yolov4x-mish.cfg/weights on COCO or your dataset by using Darknet with these modifications for each [yolo] layer in yolov4x-mish.cfg

objectness_smooth=0

scale_x_y = 1.05

remove iou_loss=ciou and iou_normalizer=0.07

as a last resort: remove new_coords=1 (and change activation=logistic to activation=linear in [convolutional] layers before [yolo] layers)

Also if you want to detect objects in crowds, then train with iou_thresh=1.0 instead of iou_thresh=0.2

SO as you said remove iou_loss=ciou , should I need to replace this loss function with any other loss??

for this readme file for detecting small objects you haven't removed 'iou_loss=ciou' https://github.com/AlexeyAB/darknet/blob/6f718c257815a984253346bba8fb7aa756c55090/cfg/yolov4.cfg#L1153

I had one question for detecting small + big objects for custom data is there any other parameters we need to consider ?? Thanks in advance. @WongKinYiu @AlexeyAB

AlexeyAB · 2021-05-23T03:02:52Z

There is fixed a bug WongKinYiu/ScaledYOLOv4#89 in Scaled-YOLOv4-CSP yolov4-csp branch for Pytorch: https://github.com/WongKinYiu/ScaledYOLOv4/tree/yolov4-csp#yolov4-csp

And performance is improved (47.8 AP -> 48.7 AP).

grandprixgp · 2021-05-25T11:24:18Z

@Hyvenos
In general, compared to yolov4 with the same resolution, the model yolov4-csp :
* increases AP (bbox coordinates) ~ +4%

* increases AP50 (number of detected objects) ~ +0.4%
  
  * increases AP50 for large objects, but **decreases AP50 for small objects**
So:
* or use yolov4-csp or yolov4x-mish with much higher resolution than yolov4

* or use yolov4
If you want to detect small objects using new models csp/x-mish, then its better to fine-tune yolov4x-mish.cfg/weights on COCO or your dataset by using Darknet with these modifications for each [yolo] layer in yolov4x-mish.cfg
* `objectness_smooth=0`

* `scale_x_y = 1.05`

* remove `iou_loss=ciou` and `iou_normalizer=0.07`

* as a last resort: remove `new_coords=1` (and change `activation=logistic` to `activation=linear` in [convolutional] layers before [yolo] layers)
Also if you want to detect objects in crowds, then train with iou_thresh=1.0 instead of iou_thresh=0.2

I hope you don't mind me asking, but what would your recommended modifications be to v4-tiny-3l that we've already modified to fit our custom data. We are streaming video and we need the low latency / high FPS hence the need for tiny, but we would like to improve:

Detection of small objects (at long distances we are facing struggles).
Flickering/blinking, which I think is due to detecting only one object at a time when multiple are present.

Our AP and AP50 values are very high, consistency and accuracy are through the roof outside of these cases (long distance and multiple objects).

Unfortunately I see most of what you've advised for small objects / multiple objects and some of the comments you've made in this thread don't apply to tiny, especially tiny-3l.

AlexeyAB · 2021-06-27T02:28:59Z

I added new models: #7414

yolov4-p5.cfg/weights
yolov4-p6.cfg/weights

This allows changing the behavior of the network while training. Setting it to 1 (disable) will have the standard YOLO behavior: match only the best anchor with a truth. However, setting it to a lower value will make any anchor with an IoU > iou_anchor_threshold with a ground truth be updated as well. See this thread for more details: AlexeyAB/darknet#7087

corentin87 · 2021-07-07T01:19:25Z

Hi everyone,

Looking at the cfg files, I can see some differences between yolov4-csp and yolov4-mish. But I can't really picture the changes. What is the main idea behind the mish compare to csp?
Thank you

WongKinYiu · 2021-07-07T02:49:31Z

main idea:
YOLOv4 = CSPDarknet + PAN
YOLOv4-CSP = CSPDarknet + CSPPAN

corentin87 · 2021-07-07T02:59:31Z

main idea:
YOLOv4 = CSPDarknet + PAN
YOLOv4-CSP = CSPDarknet + CSPPAN

Thanks but I'm asking about yolov4-csp vs yolov4-mish, not yolov4 vs yolov4-csp.

WongKinYiu · 2021-07-07T03:04:23Z

YOLOv4 = CSPDarknet(Mish) + PAN(LReLU)
YOLOv4-Mish = CSPDarknet(Mish) + PAN(Mish)
YOLOv4-CSP = CSPDarknet(Mish) + CSPPAN(Mish)

corentin87 · 2021-07-08T04:59:14Z

Thank you!

a0917bc · 2021-08-18T18:45:47Z

@AlexeyAB could we have weights and updated cfg for yolov4-tiny-3l.cfg as well?

Paper reports 28.7% AP (+6% to tiny) with same FPS on TX2.

Hello, where is the paper?
Thank you!

yusiyoh · 2021-08-20T11:00:05Z

@Hyvenos
In general, compared to yolov4 with the same resolution, the model yolov4-csp :

increases AP (bbox coordinates) ~ +4%

increases AP50 (number of detected objects) ~ +0.4%

increases AP50 for large objects, but decreases AP50 for small objects

So:

or use yolov4-csp or yolov4x-mish with much higher resolution than yolov4

or use yolov4

If you want to detect small objects using new models csp/x-mish, then its better to fine-tune yolov4x-mish.cfg/weights on COCO or your dataset by using Darknet with these modifications for each [yolo] layer in yolov4x-mish.cfg

objectness_smooth=0

scale_x_y = 1.05

remove iou_loss=ciou and iou_normalizer=0.07

as a last resort: remove new_coords=1 (and change activation=logistic to activation=linear in [convolutional] layers before [yolo] layers)

Also if you want to detect objects in crowset stride=4 instead of https://github.com/AlexeyAB/darknet/blob/6f718c257815a984253346bba8fb7aa756c55090/cfg/yolov4.cfg#L989ds, then train with iou_thresh=1.0 instead of iou_thresh=0.2

HI @AlexeyAB, For improving the detection of small objects on yolov4, you mentioned 3 changes for normal Yolov4 cfg file on the readme section i-e:
for training for small objects (smaller than 16x16 after the image is resized to 416x416) - set layers = 23 instead of

darknet/cfg/yolov4.cfg

Line 895 in 6f718c2

layers = 54

set stride=4 instead of

darknet/cfg/yolov4.cfg

Line 892 in 6f718c2

stride=2

set stride=4 instead of

darknet/cfg/yolov4.cfg

Line 989 in 6f718c2

stride=2

I tried it in past for yolov4.cfg, it results in some better outcomes, but when I did some same changes in yolov4-CSP.cfg, it's giving an error while loading network at training time, seems I did a mistake .... input shape mismatch at some layer, Can you please suggest, if we want to make same changes for Scaled yolov4 (CSP) then what exactly changes I have to do? Thanks in advance.

Hello, did you find the solution to your question? I am in the same situation, but using pytorch.

1319303003 · 2021-12-16T07:15:25Z

Dear AlexeyAB, is the pre training models of scale-yolov4 models, such as yolov4-p5 and yolov4-tiny, trained in Imagenet2012? And, in your scale-YOLOv4 paper, you said that did not use ImageNet pre-trained models. So, I want to know what dataset do you use in pre training models of yolov4-tiny? Hope to get your letter！

AlexeyAB added the enhancement label Dec 7, 2020

AlexeyAB pinned this issue Dec 7, 2020

tdurand mentioned this issue Dec 8, 2020

YOLOv4 improvements "Scaled YOLO v4" opendatacam/opendatacam#322

Open

AlexeyAB mentioned this issue Dec 15, 2020

opencv dnn module suport Scaled yolov4 model? opencv/opencv#18975

Closed

AlexeyAB mentioned this issue Dec 15, 2020

Scaled-YOLOv4 [feature request] hunglc007/tensorflow-yolov4-tflite#280

Open

jkjung-avt mentioned this issue Dec 25, 2020

YOLOv4-CSP and YOLOv4x-mish are worse than YOLOv4? jkjung-avt/tensorrt_demos#308

Closed

AlexeyAB mentioned this issue Jan 16, 2021

How do you do transfer learning based on the models from the model zoo? #7254

Open

kendyChina mentioned this issue Apr 1, 2021

When training yolov4-csp, iou_loss is a negative number. #7573

Open

casper-hansen mentioned this issue Apr 16, 2021

Table 13: YOLOv4-tiny (3l) model WongKinYiu/ScaledYOLOv4#25

Open

bjajoh mentioned this issue Jun 15, 2021

v4-CSP implemented in darknet is way worse then the normal v4 #7800

Closed

arrufat mentioned this issue Jul 13, 2021

YOLO loss davisking/dlib#2376

Merged

yusiyoh mentioned this issue Aug 3, 2021

Small object detection with Yolov4-CSP #7952

Open

kadirnar mentioned this issue Jan 10, 2022

Does it have support on the Scaled-yolov4 model? eriklindernoren/PyTorch-YOLOv3#765

Closed

AlexeyAB unpinned this issue Nov 21, 2022

Scaled-YOLOv4: Scaling Cross Stage Partial Network #7087

Scaled-YOLOv4: Scaling Cross Stage Partial Network #7087

Comments

AlexeyAB commented Dec 7, 2020 • edited Loading

MKiremitci commented Dec 7, 2020

AlexeyAB commented Dec 7, 2020

MKiremitci commented Dec 7, 2020

marvision-ai commented Dec 7, 2020 • edited Loading

tdurand commented Dec 8, 2020

philipp-schmidt commented Dec 9, 2020

marvision-ai commented Dec 9, 2020

marvision-ai commented Dec 14, 2020 • edited Loading

AlexeyAB commented Dec 15, 2020

AlexeyAB commented Dec 15, 2020

marvision-ai commented Dec 15, 2020

philipp-schmidt commented Dec 15, 2020 • edited Loading

AlexeyAB commented Dec 15, 2020

marvision-ai commented Dec 15, 2020 • edited Loading

scianand commented Dec 19, 2020

bulatnv commented Dec 22, 2020

Hyvenos commented Dec 23, 2020 • edited Loading

AlexeyAB commented Dec 24, 2020

Hyvenos commented Dec 24, 2020

AlexeyAB commented Dec 24, 2020 • edited Loading

GeekAlexis commented Jan 4, 2021

AlexeyAB commented Jan 4, 2021

stephanecharette commented Jan 4, 2021

GeekAlexis commented Jan 4, 2021 • edited Loading

AlexeyAB commented Jan 5, 2021

manoj-8246 commented Jan 6, 2021

AlexeyAB commented Jan 6, 2021

haileykwak commented Jan 20, 2021

RasmusToivanen commented Jan 22, 2021

AlexeyAB commented Jan 22, 2021

sharoseali commented Feb 2, 2021

xjinai commented Feb 17, 2021 • edited Loading

SpongeBab commented Mar 22, 2021

akashAD98 commented Apr 21, 2021 • edited Loading

AlexeyAB commented May 23, 2021

grandprixgp commented May 25, 2021

AlexeyAB commented Jun 27, 2021

corentin87 commented Jul 7, 2021

WongKinYiu commented Jul 7, 2021

corentin87 commented Jul 7, 2021

WongKinYiu commented Jul 7, 2021

corentin87 commented Jul 8, 2021

a0917bc commented Aug 18, 2021

yusiyoh commented Aug 20, 2021

1319303003 commented Dec 16, 2021

AlexeyAB commented Dec 7, 2020 •

edited

Loading

marvision-ai commented Dec 7, 2020 •

edited

Loading

marvision-ai commented Dec 14, 2020 •

edited

Loading

philipp-schmidt commented Dec 15, 2020 •

edited

Loading

marvision-ai commented Dec 15, 2020 •

edited

Loading

Hyvenos commented Dec 23, 2020 •

edited

Loading

AlexeyAB commented Dec 24, 2020 •

edited

Loading

GeekAlexis commented Jan 4, 2021 •

edited

Loading

xjinai commented Feb 17, 2021 •

edited

Loading

akashAD98 commented Apr 21, 2021 •

edited

Loading