Some of the highlights.
▶ Berkeley (and Stanford) won the best paper award: Taskonomy, a live system that quantifies how different tasks are connected and explores model-sharing and task-transfer to enable more efficient learning (a fun fact is related to this at the end).
▶ People are getting more capable of training GANs, so many papers have it as part of their training objective, and this paper claims that GAN is the correct way to optimize upon both perceptual quality and distortion metric (e.g. SSIM, PSNR).
▶ More zero-shot learning papers are popping up in the vision community, there is one for super-resolution that's particularly interesting.
▶ Michael Brown has an interesting talk title named "Welcome 大家", yes, the actual Chinese characters. He talked about harmoization between eastern and western vision community; he suggested the western commnuity paying more attention to this geographic balance and encouraged the eastern community to be more proactive.
▶ David Forsyth warns the industry to stop hiring senior vision researchers, with "dammit!!!" in italic, bold, red.
A lot of thoughts are floating in my mind. I haven't sorted them out all yet. Following are the three clearer ones (in random order).
▷ Do impactful work. I know it sounds a bit vague, but when you actually feel it, it becomes more concrete. Impactful works are those that stimulate interests, excitement and further curiosity. It’s a bit unfortunate that application-driven projects, the ones I'm more interested in, can be hard to trigger follow-up works or fundamental contribution, and that you have to find a good balance between research and product. The vision community is growing at an exponential rate, if the amount of impactful works grows accordingly, this ecosystem would be in a good shape, otherwise it's going to be more chaotic.
▷ Get involved in the community. I really like how Michael Brown brought up eastern-western harmonization, a topic that’s hardly addressed before. One of the advices he gave to the eastern world is to volunteer and to proactively get more involved. This inspires me to think out of the 'selfness' in research, but more about how I could make a contribution to this community, especially the Asian Women community. I plan to at some point volunteer to organize a workshop at a conference. I don’t feel like organizing a workshop that explicitly targets the minority, but instead a regular, technical workshop like all others. When people cannot spot or feel anything different about an asian women researcher organizing a workshop, that seems a step towards gender equality in this community.
▷ Read, read and read. Vladlen and Jitendra both brought up this in their talks. Right now reading becomes a bit more challenging in this era with a lot of hypes. You have to not only learn how to read efficiently, but also how to recognize good and bad papers with the minimum time spent. To me, my strategy would be somehow brute-force. I would choose to read with large quantity, given that I haven’t built up a good enough classifier for good and bad papers.
As part of my plan, I talked to a lot of Chinese companies and industrial labs, Alibaba, Youtu lab, Face++, Sensetime, Bytedance, etc. One reason is that I knew too little about them, and I’m still open to the career option to join a Chinese company after graduation. Another reason is that they are actually doing great and keep surprising the community with high quality commercial products and systems. I like how these Chinese companies can accurately spot the consumer demands and build robust systems from state-of-the-art research advances. However, in terms of research and originality, we are still lagging behind. Those novel and groundbreaking ideas still mostly come from the western community. Although I study in the US and represent a US institution whenever I attend a conference and give a presentation in the public, I always feel part of the eastern community, at least from my appearance and ethnicity. I really care about how we stand in the big scope, and I hope we can gain prestige and respect. Talking about that, a minor point I noticed is that the eastern community should improve on their English. It’s a bit awkward when the spotlight presenter was not able to answer a question due to language barrier.
I tried to reach out to people as much as possible. Sometimes you may admire a person’s work, but there’s not much you can chat with them, other than telling them “I really like your work on xxx.” I’m not a person good at socializing, which I should improve upon. To me, I’d choose to first build a better self through getting out good works, then I'd be more confident in having meaningful and fruitful conversations with the people I admire.
My poster session lasted a little over the supposedly two and a half hours and I didn’t have any single minute break in between. I’m satisfied with my presentation, but not that much with the project itself, considering its contribution and impact. Ren is very considerate of taking me to a cafe place to relax a bit after the poster session. Unfortunately, I didn’t get to tour around the city or visit any national park, mainly due to the time constraints and that I didn’t rent a car (plus it’s also really exhausting to meet people)! I wandered around the downtown and took some pictures of SLC. It's a clean and calm city, and unique for its background mountain view (the real mountain! not our mountain view). Overall it is a super informative experience attending this year's CVPR, and luckily I got to catch up with old friends and meet some inspiring new people. I got a better scope of what the entire field looks like and what other people are working on. There are spaces I think would be worth exploring, but also there’s the string of concentrating and focusing.
A few fun facts throughout these days:
☺ Jitendra tells people to stop writing starting sentence like "object detection is an important task in computer vision".
☺ A big screen plays World Cup live, and is super packed during the breaks (and also non-breaks).
☺ The food is really not bad, and there is ice cream for snack.
☺ More than half of the industrial sponsors are Chinese companies.
☺ The best paper requires GPU training of 40K+ hours.