LPCVC-2019

2nd Prize in Object Detection Track of 2019 IEEE Low-Power Image Recognition Challenge (LPIRC) Online Track by Orange-Control, implemented by Tensorflow (TF>=1.12).

Installation

The installation requirements are basically the same as that of the offical object-detection API of Tensorflow, please refer to Installation.md.

Dataset Preparation

1. Download [COCO-2017 dataset](http://cocodataset.org/#download) for Object Detection Task, in ./data/coco-2017.

2. Run genMinval8000.py in ./data/coco/coco-ovic to generate annotation files (i.e., ./annotations/instances_train_L.json and ./annotations/instances_minval_L.json) for training and validataion according to ovic_val_2017_list.txt required by [OVIC](https://github.com/tensorflow/models/blob/master/research/object_detection/data/mscoco_minival_ids.txt).

3. Run splitImage.py in ./data/coco/coco-ovic to split images according to ovic_val_2017_list.txt. The splitted images are stored in ./data/coco/coco-ovic/val_ovic and ./data/coco/coco-ovic/train_ovic.

4. Run removeSmallObj.py in ./data/coco/coco-ovic to generate annotation file (i.e., ./annotations/instances_train_L_256.json) for images with area larger or equal to 256 for the removal of small objects.

Training

Stage1: Load pretrained model (ssd_mobilenet_v2_quantized_coco) and train for 200K with lr 0.0004 with input size 320x320 --> AP 20.7

CUDA_VISIBLE_DEVICES=0 python ./LPCVC-2019/model_main.py \
    --pipeline_config_path=./LPCVC-2019/samples/configs/ssd_mobilenet_v2_quantized_300x300_coco_check_9.config \
    --model_dir=./LPCVC-2019/tmp/model_trained/lpcvc4_check_9/ \
    --num_train_steps=200000 \
    --sample_1_of_n_eval_examples=8 \
    --alsologtostderr

Stage2: Load model from Stage1 and train for 70K with lr 0.00005 --> AP 21.1

CUDA_VISIBLE_DEVICES=0 python ./LPCVC-2019/model_main.py \
    --pipeline_config_path=./LPCVC-2019/samples/configs/ssd_mobilenet_v2_quantized_300x300_coco_check_19.config \
    --model_dir=./LPCVC-2019/tmp/model_trained/lpcvc4_check_19/ \
    --num_train_steps=1000000 \
    --sample_1_of_n_eval_examples=8 \
    --alsologtostderr

Stage3: Load model from Stage2, change min_scale of anchor_generator to 0.15, and train with instances_train_L_256.json. Train for 150K with lr 0.0004 (AP 21.1), for 150K with lr 0.00005 (AP 21.6), for 250K with lr 0.00001 --> AP 21.7

CUDA_VISIBLE_DEVICES=0 python ./LPCVC-2019/model_main.py \
    --pipeline_config_path=./LPCVC-2019/samples/configs/ssd_mobilenet_v2_quantized_300x300_coco_check_43.config \
    --model_dir=./LPCVC-2019/tmp/model_trained/lpcvc4_check_43/ \
    --num_train_steps=1000000 \
    --sample_1_of_n_eval_examples=8 \
    --alsologtostderr

Testing

Convert .ckpt (refer to the link) to .pb for tflite

python ./LPCVC-2019/export_tflite_ssd_graph.py \
    --pipeline_config_path=./LPCVC-2019/samples/configs/ssd_mobilenet_v2_quantized_300x300_coco_check_43.config \
    --trained_checkpoint_prefix=./LPCVC-2019/tmp/model_trained/lpcvc4_check_43/model.ckpt-560197 \
    --output_directory=./LPCVC-2019/tmp/model_trained/lpcvc4_check_43/tflite/560197/non_34_convert/ \
    --max_detections=100 \
    --add_postprocessing_op=true

Convert .pb to .tflite (refer to the link)

tflite_convert --graph_def_file=./LPCVC-2019/tmp/model_trained/lpcvc4_check_43/tflite/560197/non_34_convert/tflite_graph.pb \
--output_format=TFLITE \
--output_file=./LPCVC-2019/tmp/model_trained/lpcvc4_check_43/tflite/560197/non_34_convert/model.tflite \
--input_shapes=1,320,320,3 \
--input_arrays=normalized_input_image_tensor \
--output_arrays='TFLite_Detection_PostProcess','TFLite_Detection_PostProcess:1','TFLite_Detection_PostProcess:2','TFLite_Detection_PostProcess:3' \
--inference_type=QUANTIZED_UINT8 \
--mean_values=128 \
--std_dev_values=128 \
--allow_custom_ops

Calculate AP of tflite locally

Run tflite_test.py in ./data/coco/coco-ovic to test on 7991 images in ./data/coco/coco-ovic/val_ovic.

Trained Models

You can download our trained models (.ckpt and .tflite) through the link. Note that the three models produce similar result on AP metric.

Ringhu/LPCVC-2019