Detection of small and large objects with different image resolutions and distances

[astronaut71]

Hi

Im trying to detect multiple small and large objects (workers on the construction field can have small size if image was taken from far distance or large if it was closed). I include two sample images just to show you the size of the objects trying to detect and how it looks like in the image.

Here is my data Set:

• 1 class
• 923 images for training and 116 images for validation
• Images resolution are different as were taken from different cameras such as: 2112 x 1584, 1920x1080, 3264x1836
• So workers are dynamic objects that can change the locations/positions (as they can stand/seat and work or can move)
• Training images are not same , changes in objects dimensions depending from which distance were taken.
• Detection images are almost the same as training images.

So I have some questions : 1- Which cfg file you recommend to use for training? 2- What is the recommended width and height to use for training and detection? 3- How to calculate correct anchors for my data set? 4- What is the num_of_clusters , final_width and final_height and how to calculate it for my data set?

Here the detected clusters when running ./darknet detector calc_anchors /home/darknet/build/darknet/x64/data/workers.data -num_of_clusters 9 -width 640 -height 640 -show

num_of_clusters = 9, width = 640, height = 640 read labels from 921 images loaded image: 921 box: 5858 all loaded.

calculating k-means++ ...

iterations = 63

avg IoU = 75.19 %

Saving anchors to the file: anchors.txt anchors = 8, 25, 16, 53, 21, 88, 31, 66, 31,125, 50,106, 47,199, 86,160, 72,294

Any help?

[AlexeyAB]

@astronaut71 Hi,

• I would recommend you to train yolov3-spp.cfg model with batch=64 subdivisions=64 width=832 height=832 random=1 if you have enough GPU-RAM and anchors calculated for 832x832 ./darknet detector calc_anchors /home/darknet/build/darknet/x64/data/workers.data -num_of_clusters 9 -width 832 -height 832 -show

• or if you don't have enough GPU-RAM or want to train faster: train yolov3-spp.cfg model with batch=64 subdivisions=64 width=640 height=640 random=1 and anchors calculated for 640x640 ./darknet detector calc_anchors /home/darknet/build/darknet/x64/data/workers.data -num_of_clusters 9 -width 640 -height 640 -show

• or if you want to use faster model (but less accuracy), train yolov3-tiny-3l.cfg model with batch=64 subdivisions=32 width=832 height=832 random=1 and anchors calculated for 832x832

[astronaut71]

I have RTX 2080 and installed CUDA 10. should I use max=200 parameter or? So other parameters in yolov3-spp.cfg should remain same?

[astronaut71]

can I also continue with the training using the last weights when its stopped and killed by segmentation fault (core damped) ?

[AlexeyAB]

@astronaut71

can I also continue with the training using the last weights when its stopped and killed by segmentation fault (core damped) ?

Yes, you can.

---

I have RTX 2080 and installed CUDA 10. should I use max=200 parameter or? So other parameters in yolov3-spp.cfg should remain same?

Yes, you can use max=200 especially in the last [yolo] layer.

All other parameters you can keep the same.

[astronaut71]

ok. I trained yolov3-spp.cfg model with batch=64 subdivisions=64 width=832 height=832 random=1. It stopped after 7500 iteration as can see the in chart

So should I continue training with the last weight before it stopped? Or train from beginning faster with 640x640?
Another thing is the avoiding overfitting. How can do that? So far I can reach max 90% maP and average loss not less then 0.9. Can I pre-trained the model and used some per-trained weights? Means before start training with the command

./darknet detector train /home/admini/darknet/data/workers.data /home/admini/darknet/cfg/yolov3-spp-832.cfg /home/admini/darknet/darknet53.conv.74 -map can use some pre-trained weighs to avoid over fitting or how? What should I do before running ./darknet detector train /home/admini/darknet/data/workers.data /home/admini/darknet/cfg/yolov3-spp-832.cfg /home/admini/darknet/darknet53.conv.74 -map ?

[AlexeyAB]

@astronaut71

ok. I trained yolov3-spp.cfg model with batch=64 subdivisions=64 width=832 height=832 random=1. It stopped after 7500 iteration as can see the in chart

Did you stop it manually? Or is it stopped automatically with some error?

---

Or train from beginning faster with 640x640?

Train it from the begining.

---

You can use pre-trained weights file darknet53.conv.74 for any your training process.

To check is there overfitting just set valid=valid.txt in your obj.data file so mAP calculation during training will be calculated for valid dataset. Then just get the weights file with the max of mAP by using ./darknet detector map... command.

[astronaut71]

it stopped automatically , was killed. So what should I do? Can continue the training with the last weight or start training from the beginning with 640x640 configuration file?

Yes, when started training i was using pre-trained weights file with the comand ./darknet detector train /home/admini/darknet/data/workers.data /home/admini/darknet/cfg/yolov3-spp-832.cfg /home/admini/darknet/darknet53.conv.74 -map

[AlexeyAB]

it stopped automatically , was killed.

Just try to update your code from github and recompile.

So what should I do? Can continue the training with the last weight or start training from the beginning with 640x640 configuration file?

Just continue training with width=832 height=832 random=1 and the last.weights file: /darknet detector train /home/admini/darknet/data/workers.data /home/admini/darknet/cfg/yolov3-spp-832.cfg /home/admini/darknet/backup/yolov3-spp-832_last.weights -map

[astronaut71]

Just try to update your code from github and recompile. What do you mean with this?

[AlexeyAB]

Download the latest version of Darknet with fixed bugs.

[astronaut71]

ah ok

[astronaut71]

is it valid.txt file in valid=valid.txt the test file or the train one? cause in obj.data file I have set like this

classes = 1
train  = /home/admini/darknet/data/train.txt
valid  = /home/admini/darknet/data/test.txt
names = /home/admini/darknet/data/workers.names
backup = /home/admini/darknet/backup

I my case valid.txt file is the train file. And I have test file too. Is it ok?

[AlexeyAB]

@astronaut71 Yes, it is OK. So training will use train.txt, and checking mAp (during training) will use test.txt

[astronaut71]

ok. great. Im still training and at steps 15620 can reach mAp of 95% but avg. loss is still not going down than 0.78. But the recall by mAp of 95% is 0.89. Which one is most relevant parameter for the accuracy? And is the training going good? Should keep doing it till let we say 30 000iterations?

[astronaut71]

sorry actually the obj.data file is

classes = 1
train  = /home/admini/darknet/darknet/data/train.txt
test  = /home/admini/darknet/darknet/data/test.txt
names = /home/admini/darknet/darknet/data/workers.names
backup = /home/admini/darknet/darknet/backup

is it ok test = /home/admini/darknet/darknet/data/test.txt or must bevalid = /home/admini/darknet/darknet/data/test.txt?

[AlexeyAB]

must be valid = /home/admini/darknet/darknet/data/test.txt

[AlexeyAB]

ok. great. Im still training and at steps 15620 can reach mAp of 95% but avg. loss is still not going down than 0.78. But the recall by mAp of 95% is 0.89. Which one is most relevant parameter for the accuracy?

mAP

And is the training going good?

Yes.

Should keep doing it till let we say 30 000iterations?

If mAP increases - then yes.

[astronaut71]

But those results were with test = /home/admini/darknet/darknet/data/test.txt. So means need to start train from the beginning with ./darknet detector train /home/admini/darknet/darknet/data/workers.data /home/admini/darknet/darknet/cfg/yolov3-spp-832.cfg /home/admini/darknet/darknet/darknet53.conv.74 -map

where in obj.data valid = /home/admini/darknet/darknet/data/test.txt. Correct? or can just continue training with last weights and just change the obj.data?

[AlexeyAB]

Just continue training from the last point. valid= affects only on mAP calculation.

[astronaut71]

yes. When continue training or even start from the beginning with the correct valid = /home/admini/darknet/darknet/data/test.txt then the mAP dropped to 56%. When start from the beginning by step 6500 the mAP is 56% and average loss is 1. So how to get mAP higher again? is it so low cause have only 120 test images?

[astronaut71]

or just continue training and that 56% of mAP will increase as it only by 6500 steps?

[alexanderfrey]

@astronaut71 Hi,

• I would recommend you to train yolov3-spp.cfg model with batch=64 subdivisions=64 width=832 height=832 random=1 if you have enough GPU-RAM and anchors calculated for 832x832 ./darknet detector calc_anchors /home/darknet/build/darknet/x64/data/workers.data -num_of_clusters 9 -width 832 -height 832 -show
• or if you don't have enough GPU-RAM or want to train faster: train yolov3-spp.cfg model with batch=64 subdivisions=64 width=640 height=640 random=1 and anchors calculated for 640x640 ./darknet detector calc_anchors /home/darknet/build/darknet/x64/data/workers.data -num_of_clusters 9 -width 640 -height 640 -show
• or if you want to use faster model (but less accuracy), train yolov3-tiny-3l.cfg model with batch=64 subdivisions=32 width=832 height=832 random=1 and anchors calculated for 832x832

@AlexeyAB Just for my understanding: Why is the yolov3_5l model not considered in this case ?

Thanks for your help !!