how to improve the performance in detecting small words?

[jycloud]

No description provided.

[eragonruan]

@jycloud some potential solutions to this problem.

• multi-scale testing, this is the most direct way.
• smaller anchor size when training
• feature fusion, since small words may disappear in feature map after several pooling op

[jycloud]

@eragonruan
1.multi-scale testing is a simple way,but the cnn model always resize input pictures to 1000x600,if I divide the pictures first,there is a probability to cut the character,it may cause a bad result.
2.The ANCHOR_SCALES == 16,In training phase, you mean I should reset it to 14 or smaller?
3.I dont think so,I use this code to detect scanning pictures of paper documents,my results shows:most(about 80%~90%) small words(about 20x450pixels)could be detected,but some(about more than 10%) could not.

[eragonruan]

@jycloud the model can take any size of input(short size should longer than 300), 600x1000 is just a default setting I set.
no. anchor size is 11-273

[jycloud]

@eragonruan
where can I set the model input size?
In generate_anchors.py,I find the anchor size setting,the small words' height is more than 11pix, I guess the key is not anchor size,the error is like this:

It seems every paper always miss detecting two or three rows.
how can I optimize ?

[eragonruan]

how many rows do you have, I only keep 1000 proposals during test, for your case, each row 20x450 needs about 30 proposals, hence if your document has more than 33 rows, the model may miss something

[jycloud]

@eragonruan where can I modify proposals parameter?

[eragonruan]

@jycloud lib/fast_rcnn/config.py#L179

[jycloud]

@eragonruan I reset __C.TEST.RPN_POST_NMS_TOP_N to 2000,but results did not change.

[jycloud]

@eragonruan I decide to retrain this model by my own dataset,hoped I can solve this problem,by the way,in training phase,the default setting is resize pictures to (1000,600) or (600,1000)?

[eragonruan]

@jycloud short side is resized to 600 pixel. afer resize, if long side is longer than 1200 pixel, resize the long side to 1200 pixel.

[jycloud]

@eragonruan Thank you very much.

[guddulrk]

@eragonruan where can I modify iterations (150000)? I am training the model on CPU, so it is taking too much time. I want to reduce the iterations..

Thanks

[eragonruan]

@guddulrk ctpn/text.yml#L10

[guddulrk]

Thanks, mate.

[guddulrk]

@eragonruan should I change stepsize and snapshot_iter? as there is SOLVER: Adam on #L10.

restore: 0
SOLVER: Adam
OHEM: False
RPN_BATCHSIZE: 300
BATCH_SIZE: 300
LOG_IMAGE_ITERS: 100
DISPLAY: 10
SNAPSHOT_ITERS: 1000
HAS_RPN: True
LEARNING_RATE: 0.00001
MOMENTUM: 0.9
GAMMA: 0.1
STEPSIZE: 70000
IMS_PER_BATCH: 1

[eragonruan]

@guddulrk use the latest code,stepsize is used to adjust lr, snapshot save the model

[guddulrk]

@eragonruan thank you so much.

[guddulrk]

@eragonruan I am training the model on new dataset but getting nan values as shown below:
speed: 17.140s / iter
iter: 3390 / 5000, total loss: nan, model loss: nan, rpn_loss_cls: nan, rpn_loss_box: nan, lr: 0.000010

can you please help me to sort out the issue?
Thanks

[eragonruan]

@guddulrk this may caused by your training data, check here for more detail.

[kyosocan]

@eragonruan Thanks for your code.
Now, I'm trying to change the anchor's width from 16 to 10, should I change the parameter _feat_stride as well?
Looking forward to your reply, thank you!

[eragonruan]

@kyosocan I don't think 10 is an option. the anchor's width is decided by the feature layer you choose.
for VGG net, conv5_3==>16, conv4_3==>8

[kyosocan]

@eragonruan 　Thank you very much for your reply. Now I see how to adjust this parameter. In addition, if
the resnet is better than VGG 16 in this case ?

[eragonruan]

@kyosocan sorry, I did not try model with resnet. but it's worth trying.