DIGITS: DetectNet + MSCOCO not working

Hey there.

I’m trying to train a DetectNet based on dusty example with MSCOCO. I developed a python script to convert COCO annotations to KITTI format and the built the dataset normally. I used dusty-nv example as basis: https://github.com/dusty-nv/jetson-inference

I adapted the network as follows:

name: “DetectNet” layer { name: “train_data” type: “Data” top: “data” data_param { backend: LMDB source: “/media/antonio/LinuxOutro/DIGITS/digits/jobs/20170522-184156-7112/train_db/features/data.mdb” batch_size: 10 } include: { phase: TRAIN } } layer { name: “train_label” type: “Data” top: “label” data_param { backend: LMDB source: “/media/antonio/LinuxOutro/DIGITS/digits/jobs/20170522-184156-7112/train_db/labels/data.mdb” batch_size: 10 } include: { phase: TRAIN } } layer { name: “val_data” type: “Data” top: “data” data_param { backend: LMDB source: “/media/antonio/LinuxOutro/DIGITS/digits/jobs/20170522-184156-7112/val_db/features/data.mdb” batch_size: 6 } include: { phase: TEST stage: “val” } } layer { name: “val_label” type: “Data” top: “label” data_param { backend: LMDB source: “/media/antonio/LinuxOutro/DIGITS/digits/jobs/20170522-184156-7112/val_db/labels/data.mdb” batch_size: 6 } include: { phase: TEST stage: “val” } } layer { name: “deploy_data” type: “Input” top: “data” input_param { shape { dim: 1 dim: 3 dim: 640 dim: 640 } } include: { phase: TEST not_stage: “val” } }

The LMDB source I pointed to the “data.mdb” files generated by Digits. I couldn’t find the lmdb files and considered that these were correct. Is it?

It took 24 hours to train and it’s not working at all with mAP zero. It doest give me any bbox even when I test with the train images: results

Here are the parameters for model training: configs

Here an example of image with the following label: sink 0 0 0 0.0 346.43 180.78 79.66 0 0 0 0 0 0 0 bowl 0 0 0 359.22 385.68 24.07 19.57 0 0 0 0 0 0 0

Edit1: I noticed that my labels are formatted as following: sink 0 0 0 X Y WIDTH HEIGHT 0 0 0 0 0 0 0 Should I format like this? sink 0 0 0 X Y X+WIDTH Y+HEIGHT 0 0 0 0 0 0 0

When I submit it to test for detecting with bboxes: zeros

Obviously I messed something up, can you guys help me out here?

About this issue

Original URL
State: open
Created 7 years ago
Reactions: 1
Comments: 16

Most upvoted comments

Nice!!! It was the problem. It’s training the two-class model right now and I’ll post the feedback here soon.

antoniodourado on May 31, 2017

I made a diff and found the following:

layer {
  name: "cvg/classifier"
  type: "Convolution"
  bottom: "pool5/drop_s1"
  top: "cvg/classifier"
  param {
    lr_mult: 1
    decay_mult: 1
  }
  param {
    lr_mult: 2
    decay_mult: 0
  }
  convolution_param {
    num_output: 1
    kernel_size: 1
    weight_filler {
      type: "xavier"
      std: 0.03
    }
    bias_filler {
      type: "constant"
      value: 0.
    }
  }
}

Above, “num_output: 1” is actually ‘2’, in original two-class prototxt. I couldn’t see any other problems.

ontheway16 on May 31, 2017