search

AlexeyAB / darknet

YOLOv4 / Scaled-YOLOv4 / YOLO - Neural Networks for Object Detection (Windows and Linux version of Darknet )
http://pjreddie.com/darknet/
Other
21.79k stars 7.96k forks source link

Official YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors #8595

Open AlexeyAB opened 2 years ago

AlexeyAB commented 2 years ago

Official YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors

YOLOv7:

YOLOv7x:

YOLOv7-tiny-leaky-relu:

Darknet cfg/weights file - currently tested for inference only:

Test FPS:

YOLOv7 is more accurate and faster than YOLOv5 by 120% FPS, than YOLOX by 180% FPS, than Dual-Swin-T by 1200% FPS, than ConvNext by 550% FPS, than SWIN-L by 500% FPS, than PPYOLOE-X by 150% FPS.

YOLOv7 surpasses all known object detectors in both speed and accuracy in the range from 5 FPS to 160 FPS and has the highest accuracy 56.8% AP among all known real-time object detectors with 30 FPS or higher on GPU V100, batch=1.

more5

image

yolov7_640_1280

stephanecharette commented 2 years ago

Several questions:

1) One of the huge advantages of Darknet/YOLOv[34] over other frameworks is how easy it is to incorporate into C++ applications. Will YOLOv7 be the same, with a v7.cfg file and libdarknet.so like usual, or is this new python repo where we can expect to find YOLOv7? 2) Will we have the 3 usual configurations available -- YOLOv7, YOLOv7-tiny-3L, and YOLOv7-tiny? 3) When can we expect the next release of Darknet?

akashAD98 commented 2 years ago

@AlexeyAB @WongKinYiu thank you so much for the amazing work! very much excited for yolov7

AlexeyAB commented 2 years ago

@stephanecharette Hi,

  1. YOLOv7 is mostly for Pytorch: https://github.com/WongKinYiu/yolov7

  2. There is at least YOLOv7-tiny for Darknet, but it was trained using Pytorch and converted to the Darknet. I didn't test training on Darknet yet, but it should work:

  1. I don't know.

darknet.exe detector test cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights -ext_output dog.jpg

predictions

Model inferernce time (without NMS)

darknet.exe detector demo cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights test.mp4 -benchmark

image

Model inferernce time (with NMS)

darknet.exe detector demo cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights test.mp4 -dont_show

image

toplinuxsir commented 2 years ago

@AlexeyAB Great work! Waiting for yolov7 , yolov7x for darknet .

stephanecharette commented 2 years ago

@AlexeyAB Does YOLOv7 have new "things" in the .cfg file that requires changes to the C/C++ darknet code to run correctly? Or does the Darknet code from today have everything needed to use the new YOLOv7 configuration files?

akashAD98 commented 2 years ago

@stephanecharette I tested yolov7-tiny.cfg on my custom data & its working fine.

cenit commented 2 years ago

@AlexeyAB Does YOLOv7 have new "things" in the .cfg file that requires changes to the C/C++ darknet code to run correctly? Or does the Darknet code from today have everything needed to use the new YOLOv7 configuration files?

it should already work, please open an issue if you find any problem

agjunyent commented 2 years ago

Is there a possibility to get yolov7-tiny.cfg in yaml format? Want to test the new repo, and I can only use the cfg for darknet and there's no support yet on the yolov7 repo

AlexeyAB commented 2 years ago

@stephanecharette

Does YOLOv7 have new "things" in the .cfg file that requires changes to the C/C++ darknet code to run correctly? Or does the Darknet code from today have everything needed to use the new YOLOv7 configuration files?

Training should work using these yolov7-tiny.conv.87 and yolov7-tiny.cfg files: https://github.com/AlexeyAB/darknet/issues/8595#issue-1296709732 ./darknet detector train cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.conv.87

But accuracy could be lower than training on Pytorch code, because not all data augmentation and label-assignment approaches are implemented in the Darknet.

AlexeyAB commented 2 years ago

@agjunyent

Is there a possibility to get yolov7-tiny.cfg in yaml format? Want to test the new repo, and I can only use the cfg for darknet and there's no support yet on the yolov7 repo

There is yolov7-tiny (leaky_relu) in yaml format: https://github.com/WongKinYiu/yolov7/blob/main/cfg/training/yolov7-tiny.yaml

You can try to train it using these hyper parameters: https://github.com/WongKinYiu/yolov7/blob/main/data/hyp.scratch.tiny.yaml

Just delete old train2017.cache and val2017.cache files (if you used old Pytorch-YOLO versions), and redownload labels for training: https://github.com/WongKinYiu/yolov7/releases/download/v0.1/coco2017labels-segments.zip

invo-mwiseman commented 2 years ago

@stephanecharette Hi,

  1. YOLOv7 is mostly for Pytorch: https://github.com/WongKinYiu/yolov7
  2. There is at least YOLOv7-tiny for Darknet, but it was trained using Pytorch and converted to the Darknet. I didn't test training on Darknet yet, but it should work:
  1. I don't know.

darknet.exe detector test cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights -ext_output dog.jpg

predictions

Model inferernce time (without NMS)

darknet.exe detector demo cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights test.mp4 -benchmark

image

Model inferernce time (with NMS)

darknet.exe detector demo cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights test.mp4 -dont_show

image

Those v7 weights when compared to a personal Yolov4-tiny set I've trained, using OpenCV 4.5.3 and CUDA 11.5 produce half the FPS.

v4-tiny: image

v7-tiny: image

Lost more than half the FPS. I've yet to test the accuracy between them, but the FPS trade off between a smaller increase in accuracy feels too much.

liujin1975060601 commented 2 years ago

It is necessary to develop a tool to convert yaml to cfg both for traing and for inference. We are hoping for yaml2cfg. By C++ or by python

liujin1975060601 commented 2 years ago

In the same batch, the performance of Darknet is better than that of pytorch, but the larger the batchsize is, the better the performance is. Darknet does not support half precision well, which makes the batchsize cannot be turned up. We have tried that 64/8=8 is the largest batch that can be supported, but yolov5 can support batch=16.

Can we save video memory in half precision and make the batchsize larger like yolov5？

prateekgml commented 2 years ago

@AlexeyAB @WongKinYiu can you please update the repo with steps-'How to train with a custom dataset' as we have in Darknet's repo?

clemtaylor commented 2 years ago

I tried out the yolov7-tiny.weights from above and got some rather broken output. I've run various versions of yolvo[34] and have done GPU months of training, so I don't think this is a newbie fail. Very strange...

dog-predictions-yolov7-tiny

./darknet detector test coco.51.data yolov7-tiny.cfg yolov7-tiny.weights -ext_output -i 0 data/dog.jpg
 CUDA-version: 11070 (11070), cuDNN: 8.4.0, CUDNN_HALF=1, GPU count: 2  
 CUDNN_HALF=1 
 OpenCV version: 4.5.3
 0 : compute_capability = 860, cudnn_half = 1, GPU: NVIDIA GeForce RTX 3090 
net.optimized_memory = 0 
mini_batch = 1, batch = 1, time_steps = 1, train = 0 
   layer   filters  size/strd(dil)      input                output
   0 Create CUDA-stream - 0 
 Create cudnn-handle 0
conv     32       3 x 3/ 2    416 x 416 x   3 ->  208 x 208 x  32 0.075 BF
... more network config output ...
data/dog.jpg: Predicted in 3.992000 milli-seconds.
boat: 100%  (left_x: -1407   top_y:  377   width: 2754   height:    0)
person: 99% (left_x: -1407   top_y:  377   width: 2754   height:    0)
airplane: 100%  (left_x: -1407   top_y:  377   width: 2754   height:    0)
truck: 54%  (left_x: -1407   top_y:  377   width: 2754   height:    0)
traffic light: 100% (left_x: -1407   top_y:  377   width: 2754   height:    0)
fire hydrant: 100%  (left_x: -1407   top_y:  377   width: 2754   height:    0)
stop sign: 100% (left_x: -1407   top_y:  377   width: 2754   height:    0)
parking meter: 100% (left_x: -1407   top_y:  377   width: 2754   height:    0)
bird: 99%   (left_x: -1407   top_y:  377   width: 2754   height:    0)
cat: 100%   (left_x: -1407   top_y:  377   width: 2754   height:    0)
horse: 100% (left_x: -1407   top_y:  377   width: 2754   height:    0)
elephant: 100%  (left_x: -1407   top_y:  377   width: 2754   height:    0)
bear: 100%  (left_x: -1407   top_y:  377   width: 2754   height:    0)
handbag: 100%   (left_x: -1407   top_y:  377   width: 2754   height:    0)
suitcase: 100%  (left_x: -1407   top_y:  377   width: 2754   height:    0)
skis: 100%  (left_x: -1407   top_y:  377   width: 2754   height:    0)
kite: 100%  (left_x: -1407   top_y:  377   width: 2754   height:    0)
.... many many detections ....
sports ball: 100%   (left_x:  798   top_y:  554   width:    0   height:    0)
baseball glove: 100%    (left_x:  798   top_y:  554   width:    0   height:    0)
skateboard: 100%    (left_x:  798   top_y:  554   width:    0   height:    0)
tennis racket: 100% (left_x:  798   top_y:  554   width:    0   height:    0)
wine glass: 100%    (left_x:  798   top_y:  554   width:    0   height:    0)
knife: 100% (left_x:  798   top_y:  554   width:    0   height:    0)
banana: 100%    (left_x:  798   top_y:  554   width:    0   height:    0)
apple: 100% (left_x:  798   top_y:  554   width:    0   height:    0)
orange: 100%    (left_x:  798   top_y:  554   width:    0   height:    0)
broccoli: 100%  (left_x:  798   top_y:  554   width:    0   height:    0)
toilet: 100%    (left_x:  798   top_y:  554   width:    0   height:    0)
laptop: 100%    (left_x:  798   top_y:  554   width:    0   height:    0)
remote: 100%    (left_x:  798   top_y:  554   width:    0   height:    0)
cell phone: 100%    (left_x:  798   top_y:  554   width:    0   height:    0)
oven: 100%  (left_x:  798   top_y:  554   width:    0   height:    0)
toaster: 100%   (left_x:  798   top_y:  554   width:    0   height:    0)
sink: 100%  (left_x:  798   top_y:  554   width:    0   height:    0)
refrigerator: 100%  (left_x:  798   top_y:  554   width:    0   height:    0)
clock: 100% (left_x:  798   top_y:  554   width:    0   height:    0)

The scores are all very high (mostly 100%) and the x/y coords and dimensions are 0,0 or something like 2754,0 or 0,498 (dog.jpg is only 768x576).

Oddly, with a rtsp source, it looked like some of the outputs where ~correct (cars with roughly correct bounding boxes, but low scores).

sha1sums: 5818a72c88dfa0f5fd5b9c9c0c15a82fdb8dc62d yolov7-tiny.cfg f3212b63af4764b67da6155f4950567f2183556b yolov7-tiny.weights 5120e1125cd7ba2684284fc0e547a7c38dcfc473 data/dog.jpg

AlexeyAB commented 2 years ago

@clemtaylor

darknet.exe detector test cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights -ext_output dog.jpg

image

image

clemtaylor commented 2 years ago

@AlexeyAB

Ah, thanks. The problem was there are two different versions of the yolov7-tiny.cfg, the one I was using (that I thought I downloaded with the weights) had an extra P5 section and some of the activation functions where linear and not logistic. The version from the git tree worked just fine.

kemics commented 2 years ago

Hi @AlexeyAB , really great work on yolov7

Do you have plans on yolov7 darknet cfg file release? Currently only yolov7-tiny is released

AdamCuellar commented 2 years ago

@AlexeyAB

Are the weights you provided for yolov7-tiny here the same as the yolov7-tiny.pt weights from the pytorch repo? I've loaded both in pytorch, the .weights using the previous pytorch yolo repos and the .pt in the yolov7 repo and they seem to have different values.

Ar-Ray-code commented 2 years ago

@AlexeyAB

darknet-yolov7 is compatible and I could porte to work with darknet_ros smoothly. Thanks :exclamation:

darknet_ros : https://github.com/Ar-Ray-code/darknet_ros_fp16

chjej202 commented 2 years ago

Hi, @AlexeyAB

I tested your uploaded yolov7-tiny weight file on pytorch and darknet with COCO val 2017. I got a quite different mAP result. (pytorch AP50: 0.528, darknet AP50: 0.494) Do you know why this difference happened?

I used ScaledYolov4 repo to directly run the same yolov7-tiny darknet weight file on pytorch.

The followings are details of each result. mAP result from pytorch (command: python test.py --img 416 --conf 0.001 --iou 0.65 --batch 32 --device 0 --cfg yolov7-tiny.cfg --weights yolov7-tiny.weights): Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.352 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.528 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.373 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.157 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.380 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.534 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.298 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.489 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.536 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.310 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.596 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.735

mAP result from darknet (command: ./darknet detector map cfg/coco_val.data cfg/yolov7-tiny.cfg yolov7-tiny.weights): for conf_thresh = 0.25, precision = 0.63, recall = 0.47, F1-score = 0.54 for conf_thresh = 0.25, TP = 17221, FP = 9965, FN = 19114, average IoU = 52.36 %

IoU threshold = 50 %, used Area-Under-Curve for each unique Recall mean average precision (mAP@0.50) = 0.494409, or 49.44 %

AlexeyAB commented 2 years ago

@chjej202 There are different ways for AP calculation (different ways for AUC calculation: continuous,interpolation, ...) and so on.

Try to use the same for both repos - or Pycocotool, or COCO-Codalab-server: https://github.com/AlexeyAB/darknet/wiki/How-to-evaluate-accuracy-and-speed-of-YOLOv4

In any case Pytorch could provide slightly better AP, because

chjej202 commented 2 years ago

@chjej202 There are different ways for AP calculation (different ways for AUC calculation: continuous,interpolation, ...) and so on.

Try to use the same for both repos - or Pycocotool, or COCO-Codalab-server: https://github.com/AlexeyAB/darknet/wiki/How-to-evaluate-accuracy-and-speed-of-YOLOv4

In any case Pytorch could provide slightly better AP, because

  • Pytorch version resizes network to the image aspect ratio
  • Darknet doesn't resize network

Hi @AlexeyAB

Thank you for your comment. I uploaded the results to COCO-Codalab-server, and I got the following results.

pytorch version: AP: 0.350 AP_50: 0.524 AP_75: 0.372 AP_small: 0.149 AP_medium: 0.379 AP_large: 0.534 AR_max_1: 0.296 AR_max_10: 0.476 AR_max_100: 0.509 AR_small: 0.265 AR_medium: 0.564 AR_large: 0.722

darknet version: AP: 0.330 AP_50: 0.496 AP_75: 0.351 AP_small: 0.151 AP_medium: 0.353 AP_large: 0.498 AR_max_1: 0.290 AR_max_10: 0.472 AR_max_100: 0.505 AR_small: 0.281 AR_medium: 0.553 AR_large: 0.703

For AP_50 value, the gap between two results is just slightly reduced, but pytorch version is much better than darknet version (pytorch: 0.524, darknet: 0.496, gap: 0.034 => 0.028)

Is there major difference between darknet and pytorch related to pre-/post-processing?

AlexeyAB commented 2 years ago

@chjej202

Is there major difference between darknet and pytorch related to pre-/post-processing?

For Inference and Validation - Yes.

wang1 commented 2 years ago

@clemtaylor

@AlexeyAB

Ah, thanks. The problem was there are two different versions of the yolov7-tiny.cfg, the one I was using (that I thought I downloaded with the weights) had an extra P5 section and some of the activation functions where linear and not logistic. The version from the git tree worked just fine.

HI, I got the same problem. Where are two different versions of the yolov7-tiny.cfg? I only got one from cfg directory in this repo. So, where is the correct .cfg file? Thanks a lot.

CAT1210 commented 2 years ago

I seem to be running into an error when the script tries to calculate the mAP:

cuDNN status Error in: file: C:/darknet-gpu/src/convolutional_kernels.cu : forward_convolutional_layer_gpu() : line: 553 : build time: Jun 1 2021 - 12:27:05

cuDNN Error: CUDNN_STATUS_BAD_PARAM

I don't run across this error when training any yolov4, only with training yolov7-tiny. Any ideas? Seems that it is choking on the convolutional_kernals.cu file at line 553, which reads: l.output_gpu));

AlexeyAB commented 2 years ago

I added yolov7.cfg/weights and yolov7x.cfg/weights: https://github.com/AlexeyAB/darknet/issues/8595#issue-1296709732

VisionEp1 commented 2 years ago
toplinuxsir commented 2 years ago

@AlexeyAB for yolov7x where is the .conv file for training ? Thanks!

AlexeyAB commented 2 years ago

@toplinuxsir I added it: https://github.com/AlexeyAB/darknet/issues/8595#issue-1296709732

YOLOv7x:

cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov7x.cfg weights: https://github.com/AlexeyAB/darknet/releases/download/yolov4/yolov7x.weights weights for fine-tuning: https://github.com/AlexeyAB/darknet/releases/download/yolov4/yolov7x.conv.147

toplinuxsir commented 2 years ago

@AlexeyAB Thanks!

xinsuinizhuan commented 2 years ago

now daknet support the yolov7_tiny yolov7 and yolov7x train the customed dateset? how about the accuracy whit he pytorch version?

bensonreed commented 2 years ago

@AlexeyAB Thanks for the great work. I tried to train my own dataset using yolov7x, but something seems to be going wrong. The Loss curve first decreased and then increased, and the obtained mAP has a large fluctuation. What is the reason? I modified the following parameters in the yolov7x.cfg file: batch=64，subdivisions=32，max_batches = 6000，steps=4800, 5400 filters=18,classes=1 in 3 yolo layers The training command line: darknet.exe detector train data/Oil_stain.data cfg/yolov7x.cfg yolov7x.conv.147 -map chart

cuekoo commented 2 years ago

I tested yolo v7 tiny with darknet and the results were fine. However, when tested with opencv (version 4.5.3), more than 7k objects are detected. Did I miss some settings?

import cv2 as cv
import numpy as np

def main():

    className = "./model/coco.names"

    img = cv.imread("data/dog.jpg")
    rows, cols = img.shape[:2]
    cfg = "yolov7-tiny.cfg"
    weights = "yolov7-tiny.weights"
    net = cv.dnn_DetectionModel(cfg, weights)
    net.setInputSize(cols, rows)

    net.setInputScale(1.0 / 255)
    net.setInputSwapRB(True)
    with open(className, 'rt') as f:
            names = f.read().rstrip('\n').split('\n')

    classes, confidences, boxes = net.detect(img, confThreshold=0.25, nmsThreshold=0.4)
    print("object count: {}".format(len(boxes)))

Output:

object count: 7678
ou525 commented 2 years ago

Thank you very much for sharing, can you provide a tool to convert the pytorch weights of the WongKinYiu/yolov7 into cfg, weight files

opentld commented 2 years ago

Official YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors

YOLOv7:

YOLOv7x:

YOLOv7-tiny-leaky-relu:

Darknet cfg/weights file - currently tested for inference only:

Test FPS:

  • without NMS: darknet.exe detector demo cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights test.mp4 -benchmark
  • with NMS: darknet.exe detector demo cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights test.mp4 -dont_show

YOLOv7 is more accurate and faster than YOLOv5 by 120% FPS, than YOLOX by 180% FPS, than Dual-Swin-T by 1200% FPS, than ConvNext by 550% FPS, than SWIN-L by 500% FPS, than PPYOLOE-X by 150% FPS.

YOLOv7 surpasses all known object detectors in both speed and accuracy in the range from 5 FPS to 160 FPS and has the highest accuracy 56.8% AP among all known real-time object detectors with 30 FPS or higher on GPU V100, batch=1.

  • YOLOv7-e6 (55.9% AP, 56 FPS V100 b=1) by +500% FPS faster than SWIN-L C-M-RCNN (53.9% AP, 9.2 FPS A100 b=1)
  • YOLOv7-e6 (55.9% AP, 56 FPS V100 b=1) by +550% FPS faster than ConvNeXt-XL C-M-RCNN (55.2% AP, 8.6 FPS A100 b=1)
  • YOLOv7-w6 (54.6% AP, 84 FPS V100 b=1) by +120% FPS faster than YOLOv5-X6-r6.1 (55.0% AP, 38 FPS V100 b=1)
  • YOLOv7-w6 (54.6% AP, 84 FPS V100 b=1) by +1200% FPS faster than Dual-Swin-T C-M-RCNN (53.6% AP, 6.5 FPS V100 b=1)
  • YOLOv7x (52.9% AP, 114 FPS V100 b=1) by +150% FPS faster than PPYOLOE-X (51.9% AP, 45 FPS V100 b=1)
  • YOLOv7 (51.2% AP, 161 FPS V100 b=1) by +180% FPS faster than YOLOX-X (51.1% AP, 58 FPS V100 b=1)

more5

image

yolov7_640_1280

I love this project, new bee!!!

microboym commented 2 years ago

Model inferernce time (with NMS)

darknet.exe detector demo cfg/coco.data cfg/yolov7-tiny.cfg yolov7-tiny.weights test.mp4 -dont_show

image

@AlexeyAB What device was the result generated on?

harikiran17 commented 1 year ago

Hey @AlexeyAB, I wanted to compare map scores between yolov4 and yolov7 on the coco2017 validation set. I calculated map for yolov4 using ./darknet detector map cfg/coco.data yolov4.cfg yolov4.weights where I downloaded the weights and the cfg file from model zoo and these are the results I got (I changed the width and height to 640 in the cfg file): image For yolov7 I ran with the pytorch repo using python test.py --data coco.yaml --img-size 640 --batch 32 --conf 0.25 --iou 0.5 --weights yolov7.pt --verbose and got the following results: image

I see a huge difference in map scores between yolov4 and yolov7. Is this expected? Or am I doing something wrong?

xiaoujun commented 1 year ago

cfg: https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov7.cfg weights: https://github.com/AlexeyAB/darknet/releases/download/yolov4/yolov7.weights weights for fine-tuning: https://github.com/AlexeyAB/darknet/releases/download/yolov4/yolov7.conv.132

those download link is not available anymore ，any one can download it？

cenit commented 1 year ago

they work for me. Is the problem real for any other people?

1027663760 commented 1 year ago

QQ：1027663760

srasilla commented 1 year ago

It is necessary to develop a tool to convert yaml to cfg both for traing and for inference. We are hoping for yaml2cfg. By C++ or by python

C++ please! ;)

jacob-m-nash commented 1 year ago

Hi @AlexeyAB,

I have a trained model I trained with a custom data set in Pytorch but now need it in Darkent format.

You mentioned

2. There is at least YOLOv7-tiny for Darknet, but it was trained using Pytorch and converted to the Darknet. I didn't test

How easy is it to convert my weights into Darkent format? If so, how? Or would it be best to train again using Darknet?

Thanks, Jacob

vsaw commented 1 year ago

I ran quick experiments to compare YOLOv4 and v7 performance. From what I can tell, v7 is slower and less accurate then v4. Anybody else see similar results?

invo-mwiseman commented 1 year ago

I ran quick experiments to compare YOLOv4 and v7 performance. From what I can tell, v7 is slower and less accurate then v4. Anybody else see similar results?

For me the performance is the same (or similar enough it's not noticeable), but the accuracy is notably worse.

I made early comparisons when it was released, and I was mocked for pointing out it was worse when tested in a real-world scenario. 🤣

Either way, don't bother with v5, v7 or v8. They all have the same problem.

stephanecharette commented 1 year ago

I ran quick experiments to compare YOLOv4 and v7 performance. From what I can tell, v7 is slower and less accurate then v4. Anybody else see similar results?

This is what I reported last year as well: https://www.youtube.com/watch?v=JSgDs0XXz8M

EyGy commented 1 year ago

I do agree with @vsaw @invo-mwiseman and @stephanecharette. Especially in industrial real-life use cases, I lost quite some time trying to upgrade to "newer" architectures. The performance of Scaled-yolov4 remains unbeaten. My best guess is that new versions like v7 are somewhat finetuned to benchmark datasets like COCO on an architectual level and not so well suited for general use-cases.

Would love to get some more scientific in depth insights into this, since those findings are counterintuitive....