Train cone object into tf obj-detection label map

[jacobsgomez]

The current label map used in our algorithm randomly classifies all cones in the dataset.

It's necessary to retrain a new model and patch the cone label into:

models/research/object-detection/data/COCO_complete_label_map.pbtxt

[genkilife]

Summarize what I am going to do:

• Create cone training images based on labelImg tool.
• Know how to restore COCO pre-trained model in tensorflow
• Train the model based on it.

[jacobsgomez]

Summarize what I am going to do:

• Create cone training images based on labelImg tool.
• Know how to restore COCO pre-trained model in tensorflow
• Train the model based on it.

Yes, I'm going to push a .zip of the dataset to our branch tonight. I'll touch base with you via slack.

[genkilife]

We have made progress on detecting cone objects. Following are the steps to reproduce after merging codes into codebase:

• python version Use anaconda python package

• switch to working directory “object_detection/models/research”

• Need to check tensor flow environment https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/installation.md

• create python executable files ./bin/protoc object_detection/protos/*.proto --python_out=.

• prepare training data and annotations put training images under “images” put annotation xml files under “annotations”

• transform images data into tensorflow format python3 object_detection/dataset_tools/create_cone_tf_record.py --label_map_path=object_detection/data/cone_label_map.pbtxt --data_dir=pwd —output_dir=pwd

• copy tf record into model’s data directory cp cone* FULL_CONE_BUCKET_SSD_MODEL/data/

• Review config file of tensor flow FULL_CONE_BUCKET_SSD_MODEL/models/model/ssd_mobilenet_v1_coco.config

• Begin training task python3 object_detection/legacy/train.py \
  --pipeline_config_path=${PIPELINE_CONFIG_PATH} \ --train_dir=${MODEL_DIR} \ --num_train_steps=${NUM_TRAIN_STEPS} \ --sample_1_of_n_eval_examples=$SAMPLE_1_OF_N_EVAL_EXAMPLES \ --alsologtostder

• Export model python3 object_detection/export_inference_graph.py \ --inut_type image_tensor \ --pipeline_config_path FULL_CONE_BUCKET_SSD_MODEL/models/model/ssd_mobilenet_v1_coco.config \ --trained_checkpoint_prefix FULL_CONE_BUCKET_SSD_MODEL/models/model/train/model.ckpt-0 \ --output_directory FULL_CONE_BUCKET_SSD_MODEL/models/export

• Use model to predict cones Switch to object_detection directory cd object_detection put target images in directory in “test_images” directories ** execute prediction script python3 object_detection_tutorial.py

• check output images in “out_images/test_images” directorypython version

[genkilife]

TODO:

• Relabel the images. Only pick up the cones in the inner track.
• Good to have an team account on HPC for training the model there.
• Can demo living stream with camera.