TF EAGER OBJECT DETECTION

0. Targets

• TensorFlow Eager Mode.
• Object detection models.

1. Architecture

• scripts:
  • generate_pascal_tf_records.py: generate tfrecords files from pascal source files.
  • train.py: train coco or pascal.
  • eval_pascal.py: eval pascal dataset.
  • label_map_src: copy from TensorFlow Object Detection API.
• object_detection/dataset:
  • utils:
    • label_map_utils.py: copy from TensorFlow Object Detection API.
    • tf_record_utils.py: utils to generate tfrecords files.
    • tf_dataset_utils.py: utils to generate tf.data.Dataset objects.
  • pascal_tf_dataset_generator.py: get training pascal tf.data.Dataset object from tfrecords files.
  • pascal_tf_dataset_local_file.py: get training pascal tf.data.Dataset by local files.
  • coco_tf_dataset_generator.py: get training coco tf.data.Dataset object.
  • eval_pascal_tf_dataset.py: get eval pascal tf.data.Dataset object.
• object_detection/evaluation:
  • detectron_pascal_evaluation_utils.py: copy from Detectron, eval pascal with local detection results.
  • pascal_eval_files_utils.py: generate local detection result files.
  • pascal_voc_map_utils.py: get pascal map results.
• object_detection/model:
  • faster_rcnn:
    • base_faster_rcnn_model.py: base class for faster rcnn.
    • vgg16_faster_rcnn.py: vgg16 faster rcnn model.
    • resnet_faster_rcnn.py: resnet faster rcnn model.
  • fpn:
    • base_fpn_model.py: base class for fpn.
    • resnet_fpn.py: resnet fpn model.
  • model_factory: factory for model creation.
  • anchor_target.py: generate anchor target for rpn training.
  • losses.py: smooth l1 loss & cross entropy loss.
  • prediction.py: generate predictions after roi head.
  • proposal_target.py: generate proposal target for roi training.
  • region_proposal.py: generate region proposals for both training & testing procedure.
  • roi_pooling.py: roi pooling results.
• object_detection/protos: protobuf source files.
  • protoc ./object_detection/protos/*.proto --python_out=./object_detection/protos/
• object_detection/utils:
  • anchor_generator.py: generate anchors.
  • bbox_np.py: cal iou, bbox range filter and bbox clip filter by np.
  • bbox_tf.py: cal iou, bbox range filter and bbox clip filter by tf.
  • bbox_transform.py: convert between bbox(xmin, ymin, xmax, ymax) and pred(tx, ty, tw, th)
  • visual_utils.py: draw bboxes in an image.
  • pytorch_to_tf.py: convert pytorch model to pickle map.

---

2. TODO

2.1. dataset

• pascal training dataset.
• pascal evaluating dataset.
• coco training dataset.
• coco evaluating dataset.

2.2. model

• faster rcnn
• fpn
• mask rcnn

2.3. training & evaluating

• use defun in all components.
• multi gpu support.

2.4. others

• BUG: after a few epochs, gpu memory will boomed twice... #issue 27288
• jupyter samples.
• add global step in restore variables.

---

3. training records

3.1. VOC Pascal 2007 trainval & test

Models	mAP
vgg16 tf-faster-rcnn(source)	0.708
vgg16 tf-faster-rcnn(load pre-trained model)	0.7106
vgg16 faster rcnn typical configs	0.6935/0.6869/0.6751
resnet50 faster rcnn typical configs	0.7294/0.7304
resnet101 faster rcnn tf-faster-rcnn(source)	0.757
resnet101 faster rcnn tf-faster-rcnn(load pre-trained model)	0.7578
resnet101 faster rcnn typical configs	0.7456/0.7303/0.7247/0.7261
resnet50 fpn FPN_Tensorflow(source)	0.7426
resnet50 fpn FPN_Tensorflow(load pre-trained model)	0.7430
resnet50 fpn typical configs	0.7465/0.7377/0.7392
resnet101 fpn FPN_Tensorflow(source)	0.7614
resnet101 fpn typical configs	0.7604/0.7618/0.7599

3.2. COCO 2014 minival

Models	mAP
vgg16 tf-faster-rcnn(source)	0.302
vgg16 tf-faster-rcnn(load pre-trained model)	0.302
resnet50 tf-faster-rcnn(source)	0.324
resnet50 tf-faster-rcnn(load pre-trained model)	0.324

---

4. 可有可无的教程……

• training on pascal voc 2007 trainval set, evaluating on pascal voc 2007 test set.
• Step 0: generate python protos by protoc ./object_detection/protos/*.proto --python_out=./object_detection/protos/ .
• Step 1: generate trainval datasets, set configs and use python scripts/generate_pascal_tf_records.py.
• Step 2: training by python scripts/train.py, get logs at /path/to/logs_dir/.
• Step 3: evaluating by python scripts/eval_pascal.py /path/to/logs_dir/ckpt.