Lars Nieradzik

You probably need a newer version of Tensorflow. I used version 1.5.

Hi, yeah I just used that config file without doing any big changes. However, I don't have anymore the exact config file. Since I last worked on that project, a...

Hi, instead of download_images.py, just use the COCO dataset. It is much smaller and for OCR you actually don't need so many images. You can directly download 5K images here:...

I generated the dataset myself by using a subtitle file (srt) and then doing manual annotation. I don't think that there are any datasets that you can download. Most papers...

Hi Srinath, example_1 and example_2 are easy to understand, but will often produce really bad results. example_3 and example_4 are much better (even for single object recognition) since they are...

Hi, if you change the constant WEIGHTS_FILE to a file that exists, it will work. Neural networks don't produce exactly the same results in each run. This is why, you...

It will work with any data set as long as you have the bounding boxes of the objects (xmin,ymin,xmax,ymax). If you have your data set in the VOC format, then...

Hi, yeah this is how I used the boxes too. However, in the original paper the authors apparently not only multiplied the boxes by 608 but also used different scaling...

Hi, let's look at the original IoU formula in the numerator (intersection): (min(x_A2, x_B2) - max(x_A1, x_B1))(min(y_A2, y_B2) - max(y_A1, y_B1)). Now we shift all boxes to the origin. Then...

Audio for 哟 is https://forvo.com/word/%E5%93%9F/#zh You can clearly hear a [j ɔ], not [w o]. Transcription from wiktionary: https://en.wiktionary.org/wiki/%E5%96%B2#Chinese Click on [More]. They transcribe the word as /jɔ⁵⁵/ Three phonemes...