Skip to content

Instantly share code, notes, and snippets.

@mocon

mocon/notes.md

Last active November 10, 2017 00:03
Show Gist options
  • Save mocon/2013904705ee71de5309a7511a33adf6 to your computer and use it in GitHub Desktop.
Save mocon/2013904705ee71de5309a7511a33adf6 to your computer and use it in GitHub Desktop.
Download ZIP
Off-site meeting notes
Raw
notes.md

Mobile-first video ads (Charlie)

  • $18B opportunity
  • Video inventory is scarce
  • Longer formats are less appealing
  • We have 70% viewability

Our units

  • In-Image Embedded Video (desktop only)
  • In-Screen Embedded Video
  • Video lightbox
  • 6-second Video Canvas
  • TBD mobile-first video

Competitor units

  • Kargo "Anchor" unit (6 seconds of video, then banner)
  • Celtra "Interactive" video unit
  • Undertone "Impact" video unit

Trends

  • Snapchat circular video
  • Uru (Cornell U. startup) computer vision - Recognizing negative space within videos
  • Celtra "Smart" video unit - Landing page
  • Facebook video - dominates video space (snap video to corner of screen while you browse)

KPIs

  • 70%+ viewability
  • 70%+ VCR
  • 0.1%+ CTR (secondary KPI, not as important as the first two above)
  • 15 seconds somewhat important, 6 seconds MORE important as trends continue to shift towards shorter attention spans & simpler content

In-Content ads (Merrill)

Our offerings:

  • In-Content Video with banner
  • In-Content Video without banner

Our older offerings:

  • Native Video

  • Native Content

  • Swipe to rotate 3D model (Dodge, Mini)

  • Sales not comfortable selling In-Content/Native/Video, since we aren't the only company doing those things

  • They can stress our competitive advantages (CV, NLP, other tech related to brand safety)

  • Premium publishers sell direct

  • 56% of video is outstream on mobile

  • $5B native spend

  • $33B display

Premium competitors ($100M+ annual revenue, focused on one product type)

  • Yieldmo (almost exclusively In-Content) - Animation on scroll, parallax scrolling background, carousel unit
  • Celtra/Kargo partnership - Items added as you scroll, fixed background revealed as you scroll, subtle moving background in otherwise static unit
  • Sharethrough - Native at scale

Auto products (David)

Where we are today

  • Seeking to obtain more Tier I campaigns (we might perform better here than within Tier II, since Tier I KPIs are more geared towards brand awareness instead of pure conversions)
  • Performing well in winning Tier II deals (metro/local dealers - 7-8 in a region)
  • Tier III (local dealers themselves) - We don't do this (too small, not enough money in it)

What they care about at Tier II level (lower-funnel conversions):

  • "Build a quote"
  • "Searching inventory"
  • "Locating a dealer"
  • "Schedule a test drive"

Things we could do today to improve Tier II performance:

  • Add lower-funnel buttons (CTAs) to our ads
  • "Estimate Payments"
  • "Find a Dealer"
  • Change the color of the vehicle by clicking on swatches in the unit
  • Add additional interactivity (click to learn more about several specific portions of the main creative)
  • Add in-unit "Dealer Locator" feature (just like our existing Store Locator), possibly searching local inventory, trying to get the Intender into the showroom (further down the funnel)
  • In-unit "Search Inventory" feature
  • "Build and Price" in-unitx

"Intender" is someone who wants to buy a car, tough to indentify, they may only intend to buy for ~30 days

Computer Vision (Cambron)

Current and near-term CV

  • Pre-existing tech related to people
  • Segmentation: Enabling AR overlays (classification at the pixel level)
  • Style transfer
  • "Hallucination"

Landmark detection

  • Input: face, output: feature contours
  • Hair/skin segmentation (legacy tech)
  • CNN (convolutional neural network) - Input: person, output: contours around masks (which pixels belong to which part)
  • Person detection (OD) - Input: image/video, output: bounding box around person(s)
  • CNN Skeletonization - Impose control points on a body of a human (joints) in tandem with segmentation (skin on arm etc...)
  • Hardest to tackle - Facial recognition/verification (is this Leonardo DiCaprio?)
  • Emotion from face
  • Head pose estimation (sunglasses on Divya, for Warby Parker)

Segmentation

  • Autonomous driving application
  • Picking out dominant color within the contour of a person - Cam experiment

Enabling AR: Optical flow

  • Estimating depth, via x pixel at x pixel at frame 1 vs x pixel at frame 2
  • Flow

Hallucinating (Generative applications)

  • Generate new faces from 10,000,000 other known faces
  • Simulate aging
  • Make this person smile
  • Make this woman into a man etc...
  • Unsupervised learning, still in it's infancy
  • If you can generate something, you can manipulate it

Style transfer

  • Turn Neil DeGrasse Tyson into a Picasso etc...

Memorability

  • From M.I.T. research
  • "Clap if you've seen a duplicate" experiment
  • "What regions of an image lend themselves to an overlay"

Ads with face detection (Waseem)

  • Based on publisher images
  • Challenge: Limited number of images eligible to use for facial recognition
  • We want to bring face detection into the ad itself
  • We want the ads to be scalable, not take forever to set up, and be modular

Direct competitors

  • OMMA Awards Finalist - Undertone "selfie" ad, "pick your shade", drag mouth/sunglasses onto image
  • Undertone is almost all takeover ads
  • Sephora app allows you to change your: lip, shadow, eyeliner, lash, cheek ("create your look")
  • Snapchat/Facebook/Instagram - "Filters" AR on top of image

Augmented photo ads (Dan McKenna)

  • Solve for the camera, add an image to a photo with realistic perspective
  • Should look organic, blend into the environment
  • No competitors already working on this (good story for Sales)
  • Difficult task for computer vision, need to detect a 3D plane we can project the ad onto
  • Air Jordan creative, placed in 3D into images of NBA game
  • Lufthansa creative on runway, next to airplane
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment