Ambarella AI Image Signal Processor Examined
Ambarella has announced what it calls a "breakthrough" neural network image signal processing, promising to deliver "a 10 to 100X improvement over state-of-the-art traditional ISPs" in low light, as well as improved color reproduction and dynamic range.
In this report, based on a call with Ambarella, we examine the new release, performance potential, availability, and competitive impact.
AISP ********
*********'* ******** *** ****** ******* ********* ***** into *********** ****** ******** ***** ***** ** ********. This *********** ****** ******* ************, *.*., the ****** *** ****** ******** ***** reduction ** *********** **** ** * real ****** ** ***, ** ** may ******* **** ********* ****** ** low ***** ***** ** ** **********.
*** ***** ***** ** ***** ********* claims ******* "*********** ***" ***, ** the *****, * ****** ******* ***.
Software *******, *** ********
********* **** **** **** ** ********** with *** ** ***** ************** ******** * ****** *** ** *******, allowing ****** ************* ** **** **** improvements ** ***** *** ******** ******. However, ********* *** *** ***** * list ** ********* **** *** ***** not ******* ** *** **** ******** models *** ** **********.
Low ***** *******, ** ******** ** ******
********* **** ***** ******* ************** ** AISP ** ******* ** *** ***** improvements, ***** **** **** *** ******* gains ***** ** ****. **** **** IPVM **** *** ****** ******** ***** be ******* *** **********/** ****** ** white ***** ************, ** ****, ****** they *** *** **** ********* ***** to ** ** ** * ******** for ****** **************.
Potential *********** ************
********* ****** ***** ** * **** using **** (*****, *****) **. *********** ISP (****) ** * *.** *** scene. *** ** ****** ********* ******** a **** ******** *****, **** **** saturated *****, *** **** ****** **** (seen ** ******* ** *** *********, top *****).
************ *** ***** ** **** **** of ** ******* ****** ***** (***** level *** *****, ****** as ****, **** ***** *** ******** significantly ******** *** ***** ******.
**** ****, **** *** ************ *****, these ****** *** *** ** ************** of ****** ***********.
AISP ************
** **** ** **** **** ** soon ** ********, ****** ******* ********* the ********** **** ****** *** ** available ***** ****, ***** *** ******* time ******** ** *** *** ******** and **** ******** *******.
Cost ******
********* ***** *** *** ******* **** will ** ** ******** ** ***** camera ******** **** ******** **** ** if ** ***** ** ****** **** chip ******. **** **** ** ******* these ******* ****** ** ****'* ********** availability.
Broad *** *** *********
* ***** ********* ********* ** **** Ambarella's ***** *** ****** **** ** video ************ ******* **** ******* ************* such ** ********, *****, ****, ****, Axis (****** **** **** ***** *** chips ** **** ** ***** *******) to *****'* ***** *** ********* (*************** ************ ***** *** ********* *** looking *** ***-** *** ************ ** ********** ** ********** ******.
***** ** ** ******** ** *********'* customers,**** * **** ********'* ****:
** ****, *********'* ************** ******** * relatively ****** *** ****, ** *** most, ************* ** ******* ***** ***-***** performance.
Competitive ***********
**** ***** **** ********** **** ********* similar ** ****** **********, ************* *********'* ** *** (***** *******). *******, ****** ***** *** ********** has **** ************* ******** ** ******* US ****.
**** *** **** ********* ***** ** for ******* ************, *** *** *** given ******** ***** ** *********. *** background, ***** ******* **** ****** ***** *******, Says ****, *****, ******, *************** ******-* **** ******* ********.
*******, ******* ******* ********** ******* ********** which ********* ****** ***** ** ****** encoding, ****** ** * **** ****** processing ****. ************* *** ************, ** was ******* ** ****** ***** ***. See******* ***** *** ***** ****** ********** Profile*** *******.
Potential ******
** ** ****** ****** **** **** learning / ****** ******** **** ******* low ***** ***********. *** **** ********* are *** **** *** *** **** of * ********** **** ************ **** Ambarella's **** ****.
**** ********
** *** **** ************** *** ******* "special" ******** ********** ** **** ******** (potentially ***) ***** *******
** *** ***** - ** ** a ****** *** ******* ** ********** averaging ** **** *********** "********" ** some ******* *******
****, *** *** ******* ** *** fact **** *** ***** **** ** details ***** ** *** ********* (*** not ****** "*********" ** * *********** image ********** ******** ******* **** *******). That's * *** ***** ** ******.
***, * ***** **'* * *******, e.g., ********** ** *** *** ***** in ****** ***********. *** *******, *** much *** *** ** ******* ** alter ****** ********.
******* *** *** * ****** ******* admitting ******** *** ***** **** **** -*** ****** ****** *********** **** *** Use ******, **** *** ******** ******
*** ******* **** ** ***** ******** be ******* ** ****** ** ***** issues **** ******* **** *** ********** illiterate.
*** ***** *** ***** ** ****: Rittenhouse ***** ***** ********* ***** **** feature | *** ********
*** ******* ** *** **** *** concept ** "**********", ****** **** ** "reliability".
** **** ** *** ********** *** be ********* *** ********, ** ** no ********* **** ****** ********* **** non-AI ******* **** ** ***** *********. The *** ** ** **** * judge ***/** **** ********** *** *** software ********* *** **********.
** ******, **'* **** ********* **** an ** ***** ***** * ***** may *** ** **** ** ******* the "***."
************** ** ********* *** *****, *** also *** **** ********** ** *** dataset **** *** ******** *** ***** is * *** *******.
*** ***** ** ********* ************ **** where ***** ************ ***** (*********, *********, networks) *** ****** ** **** ****** with **/**.
"************* ** ********" *** ********* **** times ** ****-*. ******** ****** ****** for ***** ****** *** ****.
*** ******* ********** ******** ****** (*.*. face ***********), ***** *** "*****" ************. Say, *** *** ********* ******** **** the ******* ** ******** ****** **** the ****** *** ******* ************* "*****" view ** *** *********.
****, ** ****** **** ******** ***** are *********** ** ******* ******* ********* in **** *********** ** * ******** frame *** ******** ** ***. *** doing ** **-***** ******** *** *** see ** **'* * ********* ** intra ***, *** ********* *** ***********. For **, *****'* *** ** **** way ** ******* **** *** *** trust *** **** ***.
****** *********** ****** ****, ***'* ******* else ******* ***** *** *** *** acceptability ** ******** ** ****** ********* with **?