chainercv | ChainerCV : a Library for Deep Learning in Computer Vision | Machine Learning library

The issue was due to a small overlooked situation where the text files had gaps as the image list was cut, copied and pasted in the same file. The text files were created in notepad. In notepad index is not visible, but the gaps are visible once you view the text files in github where the initial indexing is still present and the indexing remains even though the list was cut down in size. E.g first a list of 182 images were created but later cut down to 170. So when we use the Dataset Creation object the code reads all the lines of the text file i.e it will read 182 instead of 170. This has affected training of the model with the dataset that was incorrectly read. A new set of text files for train, val and test was created and now the training proceeded correctly.

self.converter assumes that the first argument of batch is composed of inputs that have the same shape. For example, if you use image dataset, all images are supposed to have the shape of (C, H, W).

So, could you check your dataset returns images of the same shape? And if your dataset has various shapes of images, how about using TransformDataset like https://github.com/chainer/chainercv/blob/df63b74ef20f9d8c830e266881e577dd05c17442/examples/faster_rcnn/train.py#L86?

Problem solved.

It is a problem of system encoding (Windows).

Change the System Encoding can solve the problem.

Reference:

• https://www.youtube.com/watch?v=JQU8vKXuHPE
• https://github.com/Microsoft/WSL/issues/2009

The method is a util function of Faster R-CNN, so I assume you understood what is the "anchor" proposed in Faster R-CNN.

• "Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks" https://arxiv.org/abs/1506.01497

base_size and anchor_scales determines the size of the anchor. For example, when base_size=16 and anchor_scales=[8, 16, 32] (and ratio=1.0), height and width of the anchor will be 16 * [8, 16, 32] = (128, 256, 512), as you expected. ratio determines the height and width aspect ratio.

(I might be wrong in below paragraph, please correct if I'm wrong.)

I think base_size need to be set as the size of the current hidden layer's scale. In the chainercv Faster R-CNN implementation, extractor's feature is fed into rpn (region proposal network) and generate_anchor_base is used in rpn. So you need to take care what is the feature of extractor's output. chainercv uses VGG16 as the feature extractor, and conv5_3 layer is used as extracted feature (see here), this layer is a place where max_pooling_2d is applied 4 times, which results 2^4=16 times smallen feature.

For the another question, I think your understanding is correct, py - h / 2 will be negative value. But this anchor_base value is just a relative value. Once anchor_base is prepared at the initialization of model (here), actual (absolute value) anchor is created in each forward call (here) in _enumerate_shifted_anchor method.

I wrote that code. The logic is not completely the same as py-faster-rcnn, but the difference is very subtle. Also, you can see more details here.

https://github.com/chainer/chainercv/tree/master/examples/faster_rcnn#difference-in-the-runtime-behaviour-from-the-original-code

The changes are necessary for consistency in the library. As a side note, I made that change after completely replicating the behavior of py-faster-rcnn using Chainer. I checked correctness by checking outputs of my own code and py-faster-rcnn to be within the margin of numerical errors.

I made custom class to over write some layer like this. you can control load timing for pretrained model using init flag