That's the title of your paper. Your NeurIPS spotlight presentation. The thing you put on your CV. Let me ask you something: at what point during the 50,000 iterations of training did you plan to mention that your "from scratch" model starts with a ResNet-50 backbone pretrained on 1.4 million images that humans spent YEARS labeling?
Oh, what's that? It's "standard practice"? It's "just initialization"?
THAT'S NOT INITIALIZATION. THAT'S YOUR ENTIRE WORLDVIEW. Those weights encode whether skies are blue, whether cars have wheels, whether faces have eyes in particular configurations. That's not a random seed. That's 2,048-dimensional semantic knowledge distilled from human annotation, and you're calling it "unsupervised" because you didn't use labels in the LAST part of training.
That's like inheriting a billion dollars, buying a lottery ticket with it, winning $20, and writing a memoir called Self-Made: How I Manifested Wealth Through Positive Visualization.
But it gets BETTER. Let's talk about your "bootstrapping from motion and geometry" approach. You're using optical flow to find correspondences, right? Cool, cool. And that optical flow comes from RAFT? Trained on FlyingChairs and Sintel? SYNTHETIC datasets created by HUMANS who decided what "realistic motion" looks like?
You didn't bootstrap SHIT. You inherited a sixty-year-old pipeline from computer vision researchers and slapped a transformer on the end. That's not scientific revolution. That's intellectual gentrification. You moved into photogrammetry's neighborhood, tripled the parameter count, renamed everything with "self-supervised" and "foundation," and started calling the same math "artisanal." The original residents are still here. They're called surveyors. They hate you.
Here's what ACTUALLY happened in your "self-supervised" training run:
Iteration 1: Load ImageNet weights. 1.4 million human labels of knowledge silently pour into your "unsupervised" model like bourbon into a flask at an AA meeting.
Iterations 2–10,000: Learn that objects moving together are probably the same thing — a prior you inherited from Gestalt psychology via whoever designed your grouping loss, which is to say, a HUMAN, making DECISIONS, about PERCEPTION.
Iterations 10,001–50,000: Overfit to the motion statistics of internet videos uploaded by humans, shot on cameras designed for human vision, compressed by codecs tuned for human perception, hosted on platforms that algorithmically SELECT for human engagement.
Congratulations. You've discovered that YouTube videos contain objects. Should we alert the Nobel committee, or do you want to run it for another 50k iterations first to discover that cats are popular?
Now let's talk about what ACTUALLY enrages me. The hallucination problem. The thing you absolutely REFUSE to engage with honestly.
Your monocular depth network outputs a depth value for every pixel. Every. Single. Pixel. At 100 meters. At a kilometer. At THE MOON if I point the camera up. It always gives you a number. Always confident. Always dense.
And here's the thing NOBODY in your subfield wants to say out loud: your model has NO principled mechanism to distinguish where measurement ends and hallucination begins. None. Zero. The network doesn't know. You don't know. Your reviewer didn't know. Your area chair didn't WANT to know.
When your model estimates depth at close range with strong texture and geometric cues, fine, it's doing something you could charitably call perception. But when it estimates depth for a featureless sky region 200 meters away? It's doing something COMPLETELY different. It's regurgitating training set statistics. It's pattern-matching against the ghostly average of every ImageNet photo that had sky in the upper third of the frame. It's performing a SÉANCE with dead data and reporting the results in scientific notation.
And it reports BOTH outputs — the geometrically grounded one and the statistical necromancy — with the same format. Same precision. Same decimal places. Same tidy confidence interval. As if they're the same KIND of thing.
That's not perception. That's not even regression. That's an ontological laundering operation. You take a learned prior with no geometric support, pass it through a network that outputs the same tensor format as a triangulated measurement, and by the time it reaches the evaluation script, the provenance has been washed clean. You're communing with the statistical spirits of ImageNet to CONJURE depth from the void, and you package it with uncertainty quantification like that makes it legitimate science.
You know what HONEST output looks like?
"UNMEASURABLE: Appearance-based estimate only. No geometric constraint available. Prior source: training distribution statistics. Confidence: EPISTEMICALLY VOID. This is not a measurement. This is a lookup table with gradient descent and delusions of grandeur."
But you can't put that in your paper. Because if you admitted that the MAJORITY of your output pixels are learned priors cosplaying as measurements, your "dense prediction" network becomes a "sparse measurement with vibes-based gap filling" network. And THAT doesn't get you a spotlight. That doesn't get you a Google residency. That gets you an awkward silence at the poster session.
Let's talk about benchmarks. The circular epistemological NIGHTMARE that is modern computer vision evaluation.
You train on Dataset A (ImageNet). You test on Dataset B (COCO). But COCO was DESIGNED by people who expected ImageNet-like models. It has 80 categories that HAPPEN to overlap with ImageNet's ontology. It uses bounding boxes because that's the format ImageNet normalized. It defines "good performance" as matching human annotations whose guidelines were written by researchers marinating in the same assumptions.
Then you improve your model with Private Dataset C — which you'll never release, which we can never audit, which MYSTERIOUSLY performs great on Dataset B — and you claim SOTA.
This isn't evaluation. This is an academic ouroboros that's been eating its own tail since 2012. You're optimizing for benchmarks designed by people who trained on the data you trained on, evaluated by metrics that presuppose the representations your architecture learns, reviewed by peers who use the same benchmarks and need the same system to keep working so THEIR papers get accepted.
That's not science. That's a cartel with LaTeX templates. You're doing the epistemology equivalent of insider trading and calling it "empirical validation."
And the ARROGANCE. The sheer civilizational-collapse-level HUBRIS to call these "foundation models."
Foundation models are supposed to be FOUNDATIONS. Things you build on. Bedrock. The thing under the building that touches GROUND.
Your model is a BAROQUE CATHEDRAL built on ImageNet's foundation, decorated with optical flow's stained glass windows, supported by transformer architecture's flying buttresses, funded by compute budgets that would make a defense contractor blush, and you're calling it the GROUND FLOOR.
You're not at the foundation. You're six floors up and telling everyone you're in the basement. You're so high up you've forgotten what bedrock LOOKS like.
Here's what an ACTUAL foundation looks like:
Start with sensor physics. What CAN you measure? What are the information-theoretic limits of your imaging geometry? At what range and resolution does your sensor stop providing geometric constraint and start providing decoration? KNOW THIS. Write it down. Tattoo it on your loss function.
Add physical constraints. Not learned ones. REAL ones. Gravity exists. Objects are mostly rigid over short timescales. Surfaces are piecewise smooth. Light travels in straight lines in homogeneous media. These aren't training signals. They're constraints with DERIVATIONS. You can PROVE them. Remember proofs? From before everything was ablation studies?
Layer in geographic priors. EXPLICITLY. WITH PROVENANCE. We have DEMs. We have building footprints. We have surveyed control points with known accuracy. USE THEM. Track where they came from. Mark their uncertainty. Stop treating prior knowledge as something shameful to hide in your weight initialization like a body under the floorboards.
Learn semantic priors LAST. After you've exhausted geometry. After you've exhausted physics. After you've squeezed every photon of actual information from your sensor configuration. What's left over — what geometry and physics CAN'T explain — THAT'S where learning earns its keep.
And at every stage, you maintain a LEDGER. This pixel's depth came from triangulation, here's the uncertainty. That pixel came from a DEM lookup, surveyed 2019, plus or minus 2 meters. This other pixel? This one's a learned prior with zero geometric support, and we're TELLING you that so you can decide whether to trust it instead of burying it in a dense tensor and pretending it's all the same.
THAT'S what "foundation" means. Bedrock you can build on because you know what it's made of and WHERE IT CAME FROM.
You want to do ACTUAL self-supervised learning? Train on ego-centric robot data with IMU, GPS, wheel odometry, and physics constraints. No ImageNet. No COCO. Start from a randomly initialized network and learn what triangulation MEANS geometrically, not what it looks like in FlyingChairs.
You want to do ACTUAL bootstrapping? Start with the only things you KNOW: your sensor specifications and the laws of physics. Build UP from there. Stop standing on Hinton's shoulders and pretending you're tall.
You want to do ACTUAL computer vision? Stop conflating "looks plausible to a human reviewer" with "is metrologically defensible." Stop treating confidence intervals as absolution for epistemological sins. Stop publishing papers that use the word "measurement" the way astrologers use the word "science."
Until then, you're not doing self-supervised learning. You're doing human-supervised learning with extra steps and a marketing rebrand.
You're not bootstrapping from the ground up. You're standing on ImageNet's corpse and calling it a ladder.
You're not building foundation models. You're building extensions on someone else's foundation and lying about the deed.
Your "self-supervised" depth is just supervised classification's ghost, haunting the latent space with ImageNet's unquiet dead.