How to impove small object detection

[ChungChinKei]

@AlexeyAB Hi,

Thank you so much for sharing this project.
I am trying to detect some small defects on Metal surface.

Some information about dataset:
Only 1 class.
Grayscale images.
200 images for training(100 images of positive data and 100 negative).
50 images for validation(all positive).
Images resolution is 1600 x 1200.
objects size is about 30 x 30.(even smaller)
yolov3-tiny_3l.cfg with generated anchors.

Positive sample

Negative sample

Could you give me some advise to impove mAP？

Some details of cfg:
[net]
batch=64
subdivisions=32
width=1024
height=768
channels=3
momentum=0.9
decay=0.0005
angle=0
saturation = 1
exposure = 1.5
hue= 0

learning_rate=0.005
burn_in=1000
max_batches = 10000
policy=steps
steps=8000,9000
scales=.1,.1

[yolo]
mask = 6,7,8
anchors = 21, 19, 29, 19, 27, 23, 30, 29, 36, 24, 48, 27, 40, 35, 77, 22, 56, 42
classes=1
num=9
jitter=.3
ignore_thresh = .7
truth_thresh = 1
random=1

[jonasmuriboe]

I would recommend using the full size YOLOv3 architecture like in yolov3.cfg, or maybe even yolov3-spp.cfg as this has an even higher mAP. You should try the two following modifications as well as I have personally had a drastic improvement in performance by applying them to a network that attempts to find very small objects, typically in the same size-range as you.

for training for small objects (smaller than 16x16 after the image is resized to 416x416) - set layers = -1, 11 instead of

darknet/cfg/yolov3.cfg

Line 720 in 6390a5a

layers = -1, 36

and set stride=4 instead of

darknet/cfg/yolov3.cfg

Line 717 in 6390a5a

stride=2

As well as this

recalculate anchors for your dataset for width and height from cfg-file: darknet.exe detector calc_anchors data/obj.data -num_of_clusters 9 -width 416 -height 416 then set the same 9 anchors in each of 3 [yolo]-layers in your cfg-file. But you should change indexes of anchors masks= for each [yolo]-layer, so that 1st-[yolo]-layer has anchors larger than 60x60, 2nd larger than 30x30, 3rd remaining. Also you should change the filters=(classes + 5)* before each [yolo]-layer. If many of the calculated anchors do not fit under the appropriate layers - then just try using all the default anchors.

[AlexeyAB]

@ChungChinKei Hi,

1. Try to train with default anchors.

2. Also you can try to train with SGDR:
   change these lines:

   darknet/cfg/yolov3-tiny_3l.cfg

   Lines 21 to 23 in 9c9344a

   	policy=steps
   	steps=180000,190000
   	scales=.1,.1

   
   to these lines:

policy=sgdr
sgdr_cycle=1000
sgdr_mult=2
steps=4000,6000,8000,9000
#scales=1, 1, 0.1, 0.1

---

Also you can try to use one of new models (without LSTM in filename - since LSTM is only for detection on video): #3114 (comment)

• tiny: https://github.com/AlexeyAB/darknet/files/3199622/yolo_v3_tiny_pan.cfg.txt
• or full: https://github.com/AlexeyAB/darknet/files/3199657/yolo_v3_spp_pan.cfg.txt

But they are supported only by this repository.

[ChungChinKei]

@AlexeyAB @jonasmuriboe

Thanks for your reply and suggestions .
But some errors occurred when I used the new PANet models.
How can I fix it?

[Look4-you]

@AlexeyAB Thanks for the great work!
I wonder Why is the calculated anchors less fit than the default anchors. I've try to set the anchors (which is calculated by the darknet.exe detector) at each mask layer(0.1.2/3.4.5/.6.7.8) to replace the default anchors sizes,but the come out was very unsatisfied.
If I do this wrong,how can I use the calculated anchors correctly?Could you please give me some advisement.(May be just replace some of the defult anchors but not all of them?)
Here is the calculated anchors pic:

[AlexeyAB]

@shadowmooon

How can I fix it?

You must use the latest versions of this repository: https://github.com/AlexeyAB/darknet

If I do this wrong,how can I use the calculated anchors correctly?Could you please give me some advisement.

You must change the masks (or you must change only some anchors): https://github.com/AlexeyAB/darknet#how-to-improve-object-detection

recalculate anchors for your dataset for width and height from cfg-file: darknet.exe detector calc_anchors data/obj.data -num_of_clusters 9 -width 416 -height 416 then set the same 9 anchors in each of 3 [yolo]-layers in your cfg-file. But you should change indexes of anchors masks= for each [yolo]-layer, so that 1st-[yolo]-layer has anchors larger than 60x60, 2nd larger than 30x30, 3rd remaining. Also you should change the filters=(classes + 5)* before each [yolo]-layer. If many of the calculated anchors do not fit under the appropriate layers - then just try using all the default anchors.

[Look4-you]

@AlexeyAB
How to change the masks.Could you take an example?

[duohappy]

@shadowmooon
refer to
#3372
#2960

[sadimoodi]

@ChungChinKei

did Alexy suggesiton work for you? i am doing something similiar to you so it would help me if you have any further suggestions,
thanks