search

obendidi / Tracking-with-darkflow

Real-time people Multitracker using YOLO v2 and deep_sort with tensorflow
GNU General Public License v3.0
523 stars 174 forks source link

Does self training of person is achieved here? else how can self training of person detection can be achieved for improved results. #45

Open nidhinkk opened 6 years ago

nidhinkk commented 6 years ago

Detection results in a new environment causes less detection of person. How can I improve the results. Is training of images from the previous results is automatically done in this project. what does build_train_op = help.build_train_op load_from_ckpt = help.load_from_ckpt means.

obendidi commented 6 years ago

there is no automatique training done in this project, only pretrained models are used here , you can improve detection by training your own models on your data, you can do that with :

nidhinkk commented 6 years ago

Thanks for the information. Can you tell me about the files which are generated on the location /darkflow/summary/train. What's the use of them.

On Mon, Mar 12, 2018 at 4:05 PM, Ouail notifications@github.com wrote:

there is no automatique training done in this project, only pretrained models are used here , you can improve detection by training your own models on your data, you can do that with :

-

Darkflow: https://github.com/thtrieu/darkflow

Deep_sort : https://github.com/nwojke/cosine_metric_learning

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/bendidi/Tracking-with-darkflow/issues/45#issuecomment-372263633, or mute the thread https://github.com/notifications/unsubscribe-auth/AdTqj8mHJqwJINvKGutAMvQKwVhVHJSgks5tdk95gaJpZM4SZdln .

-- Regards, Nidhin K Kurian Vuelogix Technologies Kakkanad, Kochi

obendidi commented 6 years ago

they are used for tensorboard, to visualise the darkflow model and the training params (they will be removed in next version i'm working on ) so you can just delete if they take too much space

nidhinkk commented 6 years ago

Thanks for the quick reply. I was working on detecting person count changes occurring in a specific room. Using gtts I have made an audio file of person count with some string to announce through a speaker. The audio playback will complete only after 4 seconds. By using threading I was continuously using the postprocess module, so if there is always changes in number of person, it will announce them one by one like in a queue. The difference in time of them makes the trouble for me. I need to announce only the nearly real time count by avoiding non-real time audios while I need to write the whole details of the selected frames in the csv file. If I am clear on the issue, can you have any suggestion for doing this.

On Mon, Mar 12, 2018 at 4:28 PM, Ouail notifications@github.com wrote:

they are used for tensorboard, to visualise the darkflow model and the training params (they will be removed in next version i'm working on ) so you can just delete if they take too much space

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/bendidi/Tracking-with-darkflow/issues/45#issuecomment-372269703, or mute the thread https://github.com/notifications/unsubscribe-auth/AdTqjyHtsL7BRd7ZRPuxFuQ11Mu4BiLyks5tdlTTgaJpZM4SZdln .

-- Regards, Nidhin K Kurian Vuelogix Technologies Kakkanad, Kochi

ss199302 commented 5 years ago

@bendidi I trained to generate a pb file in darkflow and it is not available in this code. That is why ah