I have implement your pmtd ideas, but the result is not good enough.

JingChaoLiu commented 5 years ago

Could you provide your implementation details of PMTD and the evaluation score of ICDAR 2017？I'm afraid that I can't help you without these details. You may need to pay attentions to these common details:

Train Stage

Train Scheduler

We train PMTD use 32 (noted as gpu_num) TITIANX 12G GPUs with SyncBatchNorm.

batch_size = gpu_num * 2 = 64
base_learning_rate = 0.00125 * batch_size = 0.08
optimizer = SGD + Multi-Step

The learning rate changes as follows:

[-8 epoch, 0 epoch) : Warmup: 8 epoch, learning rate increases exponentially from 0.00125 to 0.08
[0 epoch, 80 epoch) : Trained with a fixed learning_rate 0.08
[80 epoch, 128 epoch): Trained with a fixed learning_rate 0.008
[128 epoch, 160 epoch): Trained with a fixed learning_rate 0.0008
Loss Design
```
Loss = RPN Loss(cls + reg) + Bounding Box Loss(cls + reg) + Mask Loss
```
note: cls=classification, reg=regression

RPN Loss and Bounding Box Loss

RPN Loss and Bounding Box Loss are same with Mask RCNN, only changing the class_num from 81 to 2

Mask Loss:

mask_loss = 5 * l1_loss(predict_tensor, target_tensor)
predict_tensor = Tensor[B=positive_bbox_num, C={bg, fg}, H=28, W=28]

note: bg=background, fg=foreground target_tensor's shape is same with the predict_tensor by setting bg channel to full zero and fg channel to pyramid label

Data augmentation, RPN Anchor and OHEM

these details should implement as the paper described. Remember to random resize the image without keeping the aspect ratio

Image shape: (H, W) -> 640 * (1+3*random_1, 1+3*random_2)

Pyramid Label generation:

Draw the pyramid label on the cropped image to check its correctness
After cropped, the text region is still a polygon, but not necessarily a quadrilateral. The point num of this polygon may varies from 3 to 8, not the constant 4.
When loaded the provided pretrained model, the mask_loss between predict_tensor and target_tensor should be around 0.08
Test Stage

just follow the release code

tsing-cv commented 5 years ago

thank you for your response, I have few ops difference with your proposal.

mask loss: binary_cross_entropy
after obtain bbox with mask, using rbox regression to get rbox. I am using tf2.0 reimplement my program, leaved much bug to address.

jylins commented 5 years ago

Hi @JingChaoLiu , I have another question. Do you use ignore data(###) in ICDAR 2017 on training?

x1,y1,x2,y2,x3,y3,x4,y4,###

JingChaoLiu commented 5 years ago

Yes, we use them. The boxes with ignore=True in ICDAR 2017 are similar to those with is_crowd=True in COCO. So we follow the settings in the Mask R-CNN: Noting intersection_area = predict_box ∩ groundtruth_ignore_box, if intersection_area / groundtruth_ignore_box > 0.5 then the predict_box is set to ignore，namly neither positive nor negative.

GarrettLee commented 5 years ago

Hi @tsing-cv , could you possibly share your implementation. I would appreciate it

tsing-cv commented 5 years ago

@GarrettLee My work fell short of what the paper claimed

GarrettLee commented 5 years ago

3. When loaded the provided pretrained model, the mask_loss between `predict_tensor` and `target_tensor` should be around 0.08

Hi @JingChaoLiu , do you mean that l1_loss(predict_tensor, target_tensor) is around 0.08 or 5 *l1_loss(predict_tensor, target_tensor) is around 0.08. Thanks in advance.

JingChaoLiu commented 5 years ago

5 * l1_loss(predict_tensor, target_tensor)

soldierofhell commented 5 years ago

Mask Loss:
mask_loss = 5 * l1_loss(predict_tensor, target_tensor)
predict_tensor = Tensor[B=positive_bbox_num, C={bg, fg}, H=28, W=28]
note: bg=background, fg=foreground target_tensor's shape is same with the predict_tensor by setting bg channel to full zero and fg channel to pyramid label

Hi @JingChaoLiu , I'm trying to train with L1 loss, but inputs quickly became zeros and mask head stops to learn. I implemented your suggestion to zero input for backgroud but it didn't help:

input = mask_logits[positive_inds, labels_pos]
target = mask_targets
input = torch.where(target>0, input, torch.zeros_like(input))
l1_loss = torch.nn.L1Loss()
mask_loss = 5*l1_loss(input, target)

As for the targets everything is ok (inspected visually). I'm basicly trainging plain maskrcnn-benchmark + pyramid labels + L1loss. Pure maskrcnn-benchmark works fine and I'm just trying to improve it.

Is there anything I should tweak to make this loss converge? Thanks in advance for any suggestions

Regards,

soldierofhell commented 5 years ago

I found I had also PMTD mask predictor "MaskRCNNC4Predictor_Upsample". When I switched back to "MaskRCNNC4Predictor" mask loss seems alright now :) Any idea why bilinear upsampling causes trouble?

JingChaoLiu commented 5 years ago

I'm basicly trainging plain maskrcnn-benchmark + pyramid labels + L1loss.

In plain maskrcnn-benchmark, the mask loss is calculated by 'binary_cross_entropy', which is implemented by 'mask.sigmoid() + dot product'. When trained with pyramid label, append mask = mask.sigmoid() explicitly after the mask prediction. Then calculate the L1 Loss between the sigmoid_mask and pyramid label.

JingChaoLiu commented 5 years ago

Maybe we append a sigmoid() function inside the MaskRCNNC4Predictor_Upsample.

soldierofhell commented 5 years ago

@JingChaoLiu, thank you for pointing this out to me. Indeed there's sigmoid() in your MaskRCNNC4Predictor_Upsample head and at this moment this is the cause of bad convergence (I mean if I add it to MaskRCNNC4Predictor head I get the same effect). On the other hand, if I train with no sigmoid (which in theory is nothing wrong?) I get really nice mask pyramids.. but.. the spread between predictions is very low, I mean all valueas are really close to 0.5 and the pyramid "generator" doesn't work even if change some "constants" inside (probably some kind of rescaling would help, but I don't want to introduce more "tweaking" and I suspect the real problem is in training). Well, I guess I will struggle with it for some more time

soldierofhell commented 5 years ago

My plan for now is to:

inspect gradients in sigmoid() version
train longer non-sigmoid() version
freeze batch norm (?)
try diffrent loss like mse

Any other suggestions? :)

kalupiu commented 4 years ago

My plan for now is to:

inspect gradients in sigmoid() version

train longer non-sigmoid() version

freeze batch norm (?)

try diffrent loss like mse

Any other suggestions? :)

Same issues. Any solutions?

congjianting commented 4 years ago

请问有可用的训练代码参考吗?

STVIR / PMTD