Unsolicited advice from a Professor, Founder, and 20-time CVPR-goer
Last week, during our weekly one-on-one research meeting, one of my new PhD students innocently asked me a question: “How should I prepare for CVPR?” You see, this would be his first visit to any academic conference let alone CVPR, the prestigious one with the highest h5 rank of any conference indexed by Google Scholar. I was caught off guard; having spent the last several years focusing on my startup company, Voxel51 in an operational capacity (read: it took over all aspects of my life), I had not onboarded a new student in quite some time. So, the natural training a student gets from a mature lab was lost.
I quickly responded with some ideas of what I thought he should do to prepare. But, I wasn’t satisfied. Some of you in the CVPR community know me as a bit of a lifehacker, constantly trying to develop systems that support extracting the most value out of time spent in any activity. So, this short article is what I’ve come up with to help make your visit to CVPR 2023 in Vancouver a success.
These four strategies capture how I’d recommend making the best of CVPR. I’ll list them here and then describe each in more detail below.
- Be selective: find a signal in the noise by being intentionally selective at what you focus on.
- Be observant: consider macro trends at the conference, make an effort to summarize them.
- Be friendly: remember that CVPR is a community gathering of people with similar interests to you.
- Be active: the conference does not end when you board the flight home; actively recount your experience.
The first conference I ever attended was SIGGRAPH 2001 in Los Angeles. I was delighted that my advisor happily supported his students to travel to their first conference without having a publication yet (a practice I do with my own students today). Not being naturally outgoing, I was nervous, really nervous. Of the more than 30,000 attendees, I only knew a couple people there, and was barely through my classes as a grad student to be mildly knowledgeable about the field. I imagine this is how many new students feel today at CVPR.
Yet, I had such a fantastic time in LA that August. I was inspired by the IBM Research Augmented Reality “Everywhere Displays” installation on the exhibit floor (yes in 2001). I recall an amazing tutorial on Kalman Filters to which I still refer back today by Gary Bishop and Greg Welch. I vividly remember certain talks like the one on representing irradiance maps by Ravi Ramamoorthi. But the thing I value most was a one-on-one conversation with David Breen, a reputable faculty member in the computer graphics community; he happily engaged with me; he listened to my thoughts about certain papers; he encouraged various angles of inquiry; but in the end, the simple fact that he was happy to engage with me was infinitely valuable. Since this first conference, I’ve attended dozens if not hundreds of conferences, each one learning a bit about the field and a bit more about the people.
I hope these four strategies help you find value and connection during CVPR this week. I’ll be there as well, stop me and say hello!
There are 2,359 papers being presented at the main part of CVPR this year, not including the dozens of workshops and tutorials. I don’t know about you, but there is no way on earth I could consume this many ideas in an effective way in just a few days. However, when I first started attending conferences, I tried to do just that. Admittedly, two decades ago, the conferences were an order of magnitude smaller, so perhaps I was not actually crazy. But, no way. I would go to the conference; sit in every talk; stand at every poster; make an effort to try to understand each of the ideas. After even just one day of this, I was incapacitated. Granted, I’m a textbook introvert who expends significant energies around people. Nonetheless, this was a bad idea at its start.
What I’ve learned to do is focus. Before the start of the conference or even before each day of the conference, I select a handful of papers, maybe 3 or 4 each day, a dozen in total perhaps. Pruning 2,359 papers down to 12 is no easy task. But, seeking perfection is not the goal: you can always read another paper after the conference. I mostly select the papers based on my current areas of interest (read: keyword search), allowing myself to add a couple outside of those areas because they might be especially exciting (read: what some of my esteemed friends recommend). Before each conference day, I review those papers to build a mental model of their key ideas. I make an effort to build a sufficient scaffolding along with a question or two that would help me flesh out that scaffolding. Then during the conference day, I focus on those papers from a technical perspective, going deep with them.
This type of selectivity has allowed me to be effective at the conference working to extract significant value about those technical ideas that are closest to my own interests, while acknowledging that my attention is a limited resource. Being selective is listed first because I think it is most important: if you fail to be selective, it’s unlikely you can engage successfully in the next two strategies effectively, yes including the one about being friendly (I find it easier to break the ice if there is a common technical ground to stand on).
There is an exception to the Be Selective. Although I advocate for being very selective when it comes to what papers to go deep with, I also maintain a habit of “walking the floor” at the exhibit. I find it useful to see how technical ideas are mapped to practical applications. This may or may not resonate with you, but I recommend it.
There are a lot of ideas at CVPR. Yet, amongst these ideas are macro-trends that seem to be driving the lion’s hare of attention. For the first few CVPRs I attended, these macro trends bounced between bags of features and probabilistic graphical models. Really?! Yes, there’s no denying we as a community tend to get pretty gung ho about the latest fad. At this CVPR, one might expect diffusion or similar generative modeling ideas to be among the most popular trends in discussion, although interestingly the CVPR papers themselves were written long before the recent splashes from LLMs. Interesting times.
This second strategy recommends making an effort to observe the trends while at the conference. These are high-level collections of similar works, solutions to a particular problem, a new dataset unlocking new directions, etc. It is not necessary (or plausible, see the Be Selective strategy) to understand each of these works in fine technical detail. And, often, I have found that these observations come in unexpected ways: sometimes, they’ll come when on the industry exhibit floor talking about the impact of some trend on a practical problem; other times they’ll come in one-on-one sidebar chats with the person you’re sitting next to after the question-period of an oral. You’ll be surprised.
Make an effort to build mental bridges from recent ideas in the community to actual papers at the conference, acknowledging that so much of our community’s output is on arXiv now that these trends are likely bigger than any one conference. Importantly, what I try to do is identify not the current biggest trends in the papers and ideas I observe being discussed, but what are the undercurrents that may likely become the trends next year or the year after that. It’s not easy, but it is important.
CVPR is people. The ideas, the papers presented are the work product of one or more individuals over weeks or months. For some this is a first paper; for others it is their hundredth. Sometimes these papers represent a second or even third revision by those people. Often these papers have limitations the individuals knowingly reserve for future work; other times these papers have flaws the individuals unwittingly canonized.
None of this is trivial. Authoring an academic paper is participating in the community of scholars. It is putting your ideas out there to influence and contribute to the conversation. Authoring a paper at CVPR is not a touchdown or a home-run; it is not the conclusion of work; it is the continuance of the conversation. And a conference is probably the best venue to actively join the conversation. I mean, where else are you going to have thousands of like-minded, like-trained people to engage?
Meet the people. Build connections. Help others. I often refer to CVPR as a sort of family reunion. The people at CVPR may likely be a part of your professional (and sometimes personal) life for decades. Build relationships. Some of the most amazing relationships I have today are the result of a willingness to connect at a conference. Oftentimes, in many of these cases, I cannot tell you when a certain relationship began. At some point, they just seem to have always existed. Start now.
As the conference nears its end, you’re feeling exhausted. I’m certainly so. Yet, as you board your flight home, it’s critical to remain active in your effort to make your attendance at CVPR a success. How do you do that? Be active about your experience at CVPR. What does that mean?
This is a different strategy than the other three. It happens after the conference. For each of the other three strategies, make an active effort to deepen what you’ve gathered. For Be Selective, I highly recommend doing something specific for each of the papers you focused on. For example, you could prepare a slide for each of the papers you focused on and give the presentation to your lab-mates. You could write up a half-page summary describing the paper in your own words. You could download the paper’s code or data and work with it, if it exists.
For Be Observant, I highly recommend revisiting the observations you made about trends. Consider the trends with your peers. Look in the literature for continued evidence of those trends. Summarize your views of those trends in slides or a narrative. Do you practice Zettelkasten? Add some cards!
For Be Friendly, remain engaged with your new (and old) connections. Share your recounting on the other strategies above for their feedback. Invite them to visit your lab. Visit their lab. I’m not joking. Do it. It’ll pay off in one way or another.
I’m looking forward to the week in Vancouver at CVPR 2023. I hope these strategies make your visit a good one. Looking back on the content above, I realize this is pretty general for many contemporary, large academic conferences. I’d love to hear from you if this has helped, or if you have other ideas on life hacking CVPR. Come meet me, and others from the Voxel51 team, at booth 1618 (plus experience all the awesomeness going on there!), or reach out to me on LinkedIn.