Confessions of a researchaholic

An interactive UI that lets users manually input data and parameters to produce results is more suitable for iterative exploration (if you can wait for a few minutes) than mass production (which take take hours if not days), so I built a scripting environment from the same codebase.
I set up an experiment to run over 20+ inputs with a few values at each of the 3 parameter dimensions, and ended up with more than 4000 results which (dawned on me afterwards) are too many to examine one by one.
So I followed the path of iterating exploration as with the UI but with the already batched results without knowing for sure if I have missed any good ones.

Maybe we also need (semi)automatic tools to help us with cherry picking.

(I need to write this down while I still feel the slight amount of excitement.)

Recently I have been involved in two projects for automatic video captioning, one for a research paper and the other for a product feature.

The research paper will be presented at UIST 2023; see this page for more details.

The product feature can be accessed via this page; if you have any feedback feel free to let me know.

The two projects share some high-level ideas (such as maintaining temporal coherence for the captions while optimizing their spatial parameters with respect to the video content), but the specific methods and implementations are quite different.
Shipping a product involves a lot of testing and tuning to ensure robust experiences for a wide range of users and use cases (often beyond what the creators can initially anticipate), while publishing a research paper often requires a lot of work in writing and presentation that can be a dominating factor in deciding its acceptance and dissemination.

I heard some PhD students making fun of the SIGGRAPH dual/conference track format of (at most) 7 pages of texts and figures plus 2 pages of figures only, thinking it is something “new”.

They are too young to realize this is not new at all, and probably very old.
Before I even started my PhD program, the proceedings (1) were physically printed and (2) had separate color pages to reduce costs.
So it was common for people to put all their color figures at separate pages in the end of the paper.

I also remember that the first few times I submitted papers to SIGGRAPH, I had to mail in 6 copies of printed papers and VHS tapes for videos early enough so that they would arrive at the conference office before the deadline (while also late enough to maximize the time I can work on the submissions).
Imagine the horrors of debugging ink-jet printers and video production software/hardware in the 1990s.
I heard the story (not confirmed) that Dennis Zorin (when he was a PhD student) and (his then adviser) Peter Shroeder had a Fedex guy waiting in their office while finalizing their submission(s).
I even hand carried all the physical materials from MSR Asia’s submission to SIGGRAPH 2006 in a large suitcase (like a smuggler) and flew from Beijing to Redmond to reduce the total shipping time.

Nowadays, we can submit everything electronically at the last minute prior to the deadline, with network bandwidth/latency/reliability being the only bottleneck.