object location and detection about mxnet HOT 13 CLOSED

Same question here.

from mxnet.

I will add feature to write handcraft kernel directly in op, and we always welcome you to send PR to contribute to MXNet

from mxnet.

@antinucleon Thanks!

from mxnet.

Fast and Faster R-CNN's changes to Caffe are all in this commit and the object detection application is here.

Fast and Faster R-CNN change set
 - smooth l1 loss
 - roi pooling
 - expose phase in pycaffe
 - dropout scaling at test time (needed for MSRA-trained ZF network)

from mxnet.

Any chance to add layers which are required by faster-rcnn in near future?

from mxnet.

are there any plans for a faster-rcnn type example in the image-classification section? This would be greatly appreciated, if possible.

from mxnet.

New operator mxnet.symbol.ROIPooling is slightly tricky. Here are some key information that are missing from documentation.

import mxnet as mx
data = mx.sym.Variable('data')
# [batch_size, channel, height, width]
rois = mx.sym.Variable('rois')
# [roi_number, 5]
# last dimension is [batch index of image, x1, y1, x2, y2]
# some convolutional layer
roi_pool = mx.sym.ROIPooling(data=data, rois=rois, pooled_size=(6, 6), spatial_scale=0.0625)
# please note that batch_size changes from batch_size to roi_number after ROI pooling.

from mxnet.

@precedenceguo How to train network with mxnet.symbol.ROIPooling? Any example?

from mxnet.

I noticed that the executor_manager.DataParallelExecutorGroup uses the same slices for both input data and labels. But in the case of fast-rcnn, during training, the input images are in shape (2,3,H,W) while the labels' are (128,) for the 128 ROIS. How can we change the code for training?

from mxnet.

Multiple devices training split data into slices for devices. In this example, each data batch has image shape (2, 3, H, W) and label shape (128, ). Splitting data batches does not compromise each batch since loader supply data and label together in each batch. Therefore no change is necessary except the number of training devices. For the training of Fast R-CNN, it is recommended to see the original caffe version for now.

from mxnet.

I'm not using multiple devices for training. I use the python api model.fit which actually calls _train_multi_device. The executor_manager inside _train_multi_device uses the same slices for both input data and labels. It works for the case where the batch sizes are the same for both input data and label. I'm looking for advice on how to change the code accordingly when input data and label have different batch sizes.

from mxnet.

There could exist some issue about varying batch size with the FeedForward API. Is there any error message or anomaly?

from mxnet.

c.f. https://github.com/dmlc/mxnet/tree/master/example/rcnn

from mxnet.