Tensorflow object detection API training data set detection image score is very low and to detect the objects,
Your training set and validation set was photographed after the pixels down, probably hundreds of several hundred pixels * this, image sizes, 300 or so in the training set, is very similar between some images, 66 copies in the validation set. Use labellmg calibration label is as follows:



With Faster - R - CNN Inception v2 prototype configuration file modification, only change the corresponding path, the other parameters are the default values,
Run a python program is min_score_thresh modification to the small 0.00000005 is close to 0% to see bounding box.
Vis_util. Visualize_boxes_and_labels_on_image_array (
Image,
Np. Squeeze (boxes),
Np. Squeeze (classes). Astype (np. Int32),
Np. Squeeze (scores),
Category_index,
Use_normalized_coordinates=True,
Line_thickness=8,
Min_score_thresh=0.0000000005)
Can determine the program execution is no problem, but the suspect could be the threshold value or image training set, because this kind of building fences are together, I can only mark label each time, the surrounding is incomplete building fences, pictured above, training the detector to detect, the results are as follows:

Should also not fitting, loss a loss of image:

Could you tell me how to modify the right can be detected from the orange plastic fence?

CodePudding user response:

Problem solving,
Prior to the pre - steeped the model code modification is wrong,
Fine_tune_checkpoint: "C:/Users/z5144967/tensorflow1/models/research/object_detection/faster_rcnn_inception_v2_coco_2018_01_28/model CKPT"
This sentence before be I commented out!!!! Must retain ah!!!!!!

CodePudding user response:

reference 1st floor lcjessicalove response:

problem solving,
Prior to the pre - steeped the model code modification is wrong,
Fine_tune_checkpoint: "C:/Users/z5144967/tensorflow1/models/research/object_detection/faster_rcnn_inception_v2_coco_2018_01_28/model CKPT"
This sentence before be I commented out!!!! Must retain ah!!!!!!

Excuse me, is in the model configuration file. The config

CodePudding user response:

reference 1st floor response:

problem solving,
Prior to the pre - steeped the model code modification is wrong,
Fine_tune_checkpoint: "C:/Users/z5144967/tensorflow1/models/research/object_detection/faster_rcnn_inception_v2_coco_2018_01_28/model CKPT"
This sentence before be I commented out!!!! Must retain ah!!!!!!

thank you! Before watching a post to say to want to put the two line out, the result is how all objects detection not to come out, also look good in your answer,