Inference Speed?

[ysfalo]

HI, thanks for you excellent work! It seems your ssd version of avod costs more time than the original avod with lower AP from the results in readme. I wonder what are the pros and cons of the SSD version. Sorry, I haven't read the whole codes in the repo. And I wonder when would you release your Master's thesis, the link in readme seems not working.

[melfm]

So avod-ssd is based on avod-fpn which has inference speed of 0.1s, avod-ssd runtime speed is 0.09s, so it is slightly faster however the speed gain is not significant. I found that most of the speed bottleneck is at the feature extractor stage which is shared between single stage and two stage architectures.

Overall I found that SSD version is more prone to False Positives since the network is classifying and regressing all the anchors directly, and localization will be slightly poorer too, if we compare avod-ssd to avod-fpn since they both share the same feature extractor architecture.

The link to my thesis will be available soon!