Vesuvius Challenge awarded $1 million in prizes to computer vision researchers who had extracted text from an ancient carbonized scroll.
Last month, Vesuvius Challenge announced the 2023 award winners, who extracted the text using AI from an ancient scroll that was carbonized by volcanic ash. The scroll was "unrolled" by creating a 3D CT scan and using AI and OCR to extract text from within. The three-step approach is:
********:******** * ** **** ** * scroll ** ******** ***** *-*** **********.
************:******* *** ******** ****** ** *** rolled ******* ** *** ** **** and **** *********, ** **********, ****.
Ink *********: identifying the inked regions in the flattened segments using a machine learning model.
********
*** ****** **** *** ***** ** the ***** ** **. *********** ******* ** ******, *** * ** ***** *** was ******* *** *********** ** ******* text. (******** ************* ** ************ * 3D ****** *****:)
************
*** ** **** *** ********* **** layers, *** ********* *** ** ** perform *** *** ******* ****. **** segment ** *** ****** *** ********* for *** ***********, *** ** ******* text.
Ink *********
******** ********* ******** *********** ********* * ******** ** ** ***** 140 ********** ****, *** * ***** ** ************* students ************ ********* **** ************* ******** *** *** ******* ** the ************ ** ***** ** ******** received.
*** ********* ********* *** ******* ******** to ***** ***** ********* ***** *************. Image ********** ******** *********** ******** * ****** ****-************ ***** *** ink *********.
*** ********** ******** ******* **** ***** different ***** *************, **** ********** *** findings ** *** ******, **** *** strongest ****** ***** ****** **** * TimeSformer-based *****. ******** ******** ******* *********** and *************, ********* ******* **** ******** architectures, * ***** ****** *****/****** ****** sizes, ***** *********, *** ******* ********** folds. **** **** *** *** ******, this *** ********* **** *** **** made ****** ** **** ****** (** GitHub), ******** ** ******** ** *** community.
** ******** ** ************ *** *********, the ******* ********** ********* *** ********* auto-segmentation ******** ** **** **** ** date (**** ***** *** ******* ** “segmentation” *****). ******************* (*******: ******* *********) by ****** ********* ******* ******* ******** from ******** *******. **-************* ** **** known ***** ******** ******** *** ********, and ******** *** ************* ****** ******* elsewhere, **** ** *** ********* **** of *** ******!
*** ***** *** *** ********** ** learning **** ***** *** ******** ********* and *** *******, **** *** **** resources:
***** **** *% ** *** ****** has **** **********, ** *** ***** researchers **** ******* *** ****? **** AI ******* ********, **** *** *** implications ** ******** ***** ********?
** ********* ********** **** *** ****** talks *****, ** *** ***** *** availability ** ***** ******* *** ******** received ** ****? (**. ********, ******, rare ****, ***.)
** ** ****** **%? **** ******** the ***** ********, ** **** *** impression ** *** ******* ** **** perfect.