centernet_tensorflow_wilderface_voc

1. Introduction

• This is the unofficial implementation of the "CenterNet:Objects as Points".In my experiment, it was not based on the DLA34, Hourglass and other networks in the original paper. I simply modified shufflenetv2_1.0x and yolov3, and kept their feature extraction part, then connected to centernet_detect_head, and did not use dcn convolution.
• This is just a simple attempt to the effect of the algorithm.I only have one 1080ti,I did not use any data augmentation and any other tricks during training，so the model is not very good,still need more work to get good results.If it helps you, please give me a star.You can read my Chinese notes.(1)https://zhuanlan.zhihu.com/p/68383078 (2)https://zhuanlan.zhihu.com/p/76378871
• Official implementation:https://github.com/xingyizhou/CenterNet
• CenterNet:Objects as Points:https://arxiv.org/pdf/1904.07850.pdf
• Shufflenetv2 is modified from:https://github.com/timctho/shufflenet-v2-tensorflow
• Shufflenetv2:https://arxiv.org/abs/1807.11164
• Yolov3 is is modified from:https://github.com/wizyoung/YOLOv3_TensorFlow
• Yolov3:https://pjreddie.com/media/files/papers/YOLOv3.pdf
• ExFuse:https://arxiv.org/abs/1804.03821

2. My experimental environment

• anaconda3、pycharm-community、python3.6、numpy1.14
• tensorflow1.13、slim
• cuda10.0、cudnn7.6
• opencv-python4.1
• gtx1080ti*1

3. datasets

• For single-target detection, trained on wilderface dataset with 12876 training images.
• For multi-target detection, trained on pascal-voc2012 dataset with 17125 training images.

4. Experimental

4.1 Modified the heat map generation method to solve the problem that the loss cannot be optimized

pic1 src_bbox_gt

pic2 heatmap_original and heatmap_modified

pic3 heatmap_original_box and heatmap_modified_box

4.2 Face detection(wilder face)

input_size:512x512
downsample_ratio:4.0
batch_size:14
global_steps:14800
epochs≈16

pic4 shufflenetv2_face_result

4.3 Multi-target detection(voc)

yolov3_centernet:
input_size:512x512
downsample_ratio:8.0
batch_size:8
global_steps:40000
epochs≈18

shufflenetv2_centernet:
input_size:512x512
downsample_ratio:4.0
batch_size:16
global_steps:40000
epochs≈37

shufflenetv2_seb_centernet:
input_size:512x512
downsample_ratio:4.0
batch_size:16
global_steps:40000
epochs≈37

4.3.1 Network

pic5 yolov3_centernet_voc

pic6 shufflenetv2_centernet_voc

pic7 shufflenetv2_centernet_seb_voc

4.3.2 result(on training set,not very good on the test set)

pic8 shufflenetv2_centernet_voc_result

4.4 tensorboard loss curve

pic9 yolov3_centernet_voc_total_loss

pic10 shufflenetv2_centernet_voc_total_loss

4.5 inference time

environment：python3.6 gtx1080ti*1 intel-i7-8700k
model_name   			avg_time(ms)    input_size	 model_size(.pb)	
shufflenetv2_face_v1	        21.37	        512x512		 20.5MB
shufflenetv2_voc_v2		17.4		512x512		 24.9MB
yolo3_voc_v2		        25.53		512x512	         227.7MB

5. Run test demo(still need more work to get good results)

download ckpt filehttps://pan.baidu.com/s/1OVtOyHdc6qgcvTn56s5m2wcode:qd35,and put them to ./shufflenetv2_face_V1/, ./shufflenetv2_seb_voc/, ./shufflenetv2_voc/,and ./yolo3_voc/,then run test_voc_on_images.py or test_face_on_images.py

6.Create tfrecords to train

• The function about how to create and parse tfrecords is under folder img2tfrecords_detection.
• You only need to modify the following variables：img_path, txt_path, tfrecords.
• Then run img2tfrecords_pad.py to create tfrecords and parse it by parse-tfrecords.py.
• For detailed implementation, please see the relevant code under folder img2tfrecords_detection.