Unable to convert Imagenet-VID to TFRecords for TF/lstm_object_detection

[evgps]

Dear @wakanda-ai and @KanaSukita I have found that you have succeeded in running code from tensorflow/models#6253 .
Can you please provide more instructions on how to use your convertor and which settings in https://github.com/tensorflow/models/blob/master/research/lstm_object_detection/configs/lstm_ssd_mobilenet_v1_imagenet.config are correct for usage with your code.

[KanaSukita]

Hi @evgps the command I use for the converter is:

python vid_2015_to_tfrecord.py --root_dir=/path/to/VID2015 --output_path=/path/to/output

It should generate tfrecords containing sequential data with length of 4, which is the same in the config. And in the config, you need to change the input path to your tfrecords.

BTW, optimization for MobilenetV2 is updated recently, you may try that. However, still I couldn't train a valid model using the newest config, so I doubt that the way I convert the tfrecord might be wrong.

[wangpichao]

@KanaSukita How to set the lab_map_path? Or what is it? How to generate it? Any help would be appreciated.
train_input_reader: {
shuffle_buffer_size: 32
queue_capacity: 12
prefetch_size: 12
min_after_dequeue: 4
label_map_path: "path/to/label_map"
external_input_reader {
[lstm_object_detection.protos.GoogleInputReader.google_input_reader] {
tf_record_video_input_reader: {
input_path: './research/lstm_object_detection/inputs/tfrecords/VID_2015-0.tfrecord'
data_type: TF_SEQUENCE_EXAMPLE
video_length: 4
}
}
}
}

[KanaSukita]

@wangpichao I write the vid_label_map.pbtxt myself, something like below:

item {
name: "n02691156"
id: 1
display_name: "airplane"
}
item {
name: "n02419796"
id: 2
display_name: "antelope"
}
item {
name: "n02131653"
id: 3
display_name: "bear"
}
...

[wangpichao]

@wangpichao I write the vid_label_map.pbtxt myself, something like below:

item {
name: "n02691156"
id: 1
display_name: "airplane"
}
item {
name: "n02419796"
id: 2
display_name: "antelope"
}
item {
name: "n02131653"
id: 3
display_name: "bear"
}
...

@KanaSukita thanks. Till now, are you still working on this project?

[KanaSukita]

@wangpichao No, I couldn't train a valid model even if I use the training data for evaluation. But I am willing to try again if there is any upate in the official repo.

[wangpichao]

@wangpichao No, I couldn't train a valid model even if I use the training data for evaluation. But I am willing to try again if there is any upate in the official repo.

@KanaSukita I also tried several times and with no valid models.