Detectron Transfer Learning with PASCAL VOC 2007 dataset

3 min readFeb 15, 2018

** Detectron implemented several object detecton algorithms. All the algorithms are trained on coco 2014 data set which has 80 categories. I want to fine tune the faster-rcnn with FPN on pascal voc 2007 dataset which has only 20 categories. The same way can be used to fine tune your own model on a new dataset **

1. Setup caffe2 and Detectron and run the Detectron demo successfully.

I will refer the Detectron directory as $DETECTRON

2. Download the pre-trained model

The code will download the models automatically. But my internet is slow and I’d like to download them before I run the code. Becase I’m gonna using the ResNet-50 as the backbone, so I need to download the ResNet and faster_rcnn_R-50-FPN model.

wget https://s3-us-west-2.amazonaws.com/detectron/ImageNetPretrained/MSRA/R-50.pkl /tmp/detectron/detectron-download-cache/ImageNetPretrained/MSRA/R-50.pkl
wget https://s3-us-west-2.amazonaws.com/detectron/36225732/12_2017_baselines/mask_rcnn_R-50-FPN_2x.yaml.08_43_08.gDqBz9zS/output/train/coco_2014_train%3Acoco_2014_valminusminival/generalized_rcnn/model_final.pkl

3. Prepare configuration file.

a. Copy the sample configure file from $DETECTRON/configs/getting_started

cd $DETECTORN
mkdir experiments && cd experiments
cp ../configs/getting_started/tutorial_1gpu_e2e_faster_rcnn_R-50-FPN.yaml e2e_faster_rcnn_resnet-50-FPN_pascal2007.yaml

b. Change the configuration file

MODEL:
  TYPE: generalized_rcnn
  CONV_BODY: FPN.add_fpn_ResNet50_conv5_body
  NUM_CLASSES: 21
  FASTER_RCNN: True

The pascal voc 2007 has only 20 classes plus one background class. So the NUM_CLASSES is set to 21.

TRAIN:
  SNAPSHOT_ITERS: 5000
  WEIGHTS: /tmp/detectron/detectron-download-cache/35857389/12_2017_baselines/e2e_faster_rcnn_R-50-FPN_2x.yaml.01_37_22.KSeq0b5q/output/train/coco_2014_train%3Acoco_2014_valminusminival/generalized_rcnn/model_final.pkl
  DATASETS: ('voc_2007_train',)

Change the WEIGHTS value to where you just place in step 2.

4. Download pascal voc2007 and coco format annotations.

Refer data readme file to prepare the pascal data set.

The code support pascal data set has bug, the code in $DETECTRON/lib/datasets/dataset_catalog.py should be changed as following:

'voc_2007_train': {
        IM_DIR:
            _DATA_DIR + '/VOC2007/JPEGImages',
        ANN_FN:
            _DATA_DIR + '/VOC2007/annotations/pascal_train2007.json',
        DEVKIT_DIR:
            _DATA_DIR + '/VOC2007/VOCdevkit2007'
    },
    'voc_2007_test': {
        IM_DIR:
            _DATA_DIR + '/VOC2007/JPEGImages',
        ANN_FN:
            _DATA_DIR + '/VOC2007/annotations/pascal_test2007.json',
        DEVKIT_DIR:
            _DATA_DIR + '/VOC2007/VOCdevkit2007'
    },

5. Change the cls_score and bbox_pred name to prevent error when train_net.py load weights

In lib/modeling/fast_rcnn_heads.py, change all cls_score to cls_score_voc, bbox_pred to bbox_pred_voc.

6. Run command to begin training.

python2 tools/train_net.py --cfg experiments/e2e_faster_rcnn_resnet-50-FPN_pascal2007.yaml  OUTPUT_DIR experiments/output

7. Copy the final model just trained.

mkdir -p /tmp/detectron-download-cache/voc2007/
cp experiments/output/train/voc_2007_train/generalized_rcnn/model_iter49999.pkl /tmp/detectron-download-cache/voc2007/model_final.pkl

8. Infer some images psacal 2007 from test dataset

python2 tools/infer_simple.py --cfg experiments/e2e_faster_rcnn_resnet-50-FPN_pascal2007.yaml \
    --output-dir /tmp/detectron-visualizations --wts /tmp/detectron-download-cache/voc2007/model_final.pkl \
    demo2

Unfortunately, I found all person is labeld as bird. This maybe caused by the json dataset is not corrrectly converted.

9. Run test_net.py on pascal 2007 test dataset.

python2 tools/test_net.py \
    --cfg experiments/e2e_faster_rcnn_resnet-50-FPN_pascal2007.yaml \
    TEST.WEIGHTS /tmp/detectron-download-cache/voc2007/model_final.pkl \
    NUM_GPUS 1

The test report the AP and mAP:

INFO voc_dataset_evaluator.py: 127: AP for aeroplane = 0.8095
INFO voc_dataset_evaluator.py: 127: AP for bicycle = 0.8042
INFO voc_dataset_evaluator.py: 127: AP for bird = 0.7086
INFO voc_dataset_evaluator.py: 127: AP for boat = 0.6418
INFO voc_dataset_evaluator.py: 127: AP for bottle = 0.6861
INFO voc_dataset_evaluator.py: 127: AP for bus = 0.8822
INFO voc_dataset_evaluator.py: 127: AP for car = 0.8794
INFO voc_dataset_evaluator.py: 127: AP for cat = 0.8621
INFO voc_dataset_evaluator.py: 127: AP for chair = 0.5876
INFO voc_dataset_evaluator.py: 127: AP for cow = 0.7799
INFO voc_dataset_evaluator.py: 127: AP for diningtable = 0.7404
INFO voc_dataset_evaluator.py: 127: AP for dog = 0.8497
INFO voc_dataset_evaluator.py: 127: AP for horse = 0.8855
INFO voc_dataset_evaluator.py: 127: AP for motorbike = 0.7912
INFO voc_dataset_evaluator.py: 127: AP for person = 0.7931
INFO voc_dataset_evaluator.py: 127: AP for pottedplant = 0.5142
INFO voc_dataset_evaluator.py: 127: AP for sheep = 0.7950
INFO voc_dataset_evaluator.py: 127: AP for sofa = 0.7457
INFO voc_dataset_evaluator.py: 127: AP for train = 0.7956
INFO voc_dataset_evaluator.py: 127: AP for tvmonitor = 0.6960
INFO voc_dataset_evaluator.py: 130: Mean AP = 0.7624