Latest / Logosofia no Dia a Dia / Change Detection AI for Construction Sites

Transcript

Change Detection AI for Construction Sites · Logosofia no Dia a Dia

• [Prompt] custom: I'd love to talk today about visual change detection systems using AI and specific models for that. | Hosts: corn, herman [Response] Corn: Daniel sent us this one — he's asking about visual change detection AI, specifically for monitoring a construction site with daily photographs. You've got the same patch of ground, same camera angle, but every day it's different trucks, different people, different excavation depth. A layperson sees noise. An expert like a civil engineer spots process stages and equipment swaps. But Daniel's real question is about that third layer — what an AI can see that both the layperson and the expert might miss, and what kind of model actually does this work. He's also curious about the platforms and tools available, and whether the same approach transfers to things like medical imaging. Herman: Right, and the construction site is such a perfect microcosm for this problem because it's messy in exactly the ways that break naive approaches. You've got lighting changes day to day, shadows moving across the pit, rain turning dirt to mud, camera vibrations shifting the frame by a few pixels. A simple pixel subtraction between two photos would just light up like a Christmas tree with false positives. Corn: The "everything changed" detector. Herman: Yet buried in all that noise, there might be a hairline crack propagating along a retaining wall, or a piece of equipment that was moved overnight without authorization. Things that even a trained engineer reviewing the photos might gloss over because they're focused on expected progress. The human brain has this fascinating filtering mechanism — we see what we expect to see. An AI doesn't have that expectation, which is both its strength and its weakness. Corn: It's equally capable of flagging a genuine structural risk and a pigeon that happened to land in frame. Herman: That's the tension. And that tension is why the model architecture actually matters — you can't just throw a generic image classifier at this and hope for the best. This is a defined subfield of computer vision called change detection, and it's been dominated for years by satellite imagery applications. Urban development tracking, deforestation monitoring, disaster assessment. But what's shifted in the last year or two is that foundation models like SAM and DINOv2 have gotten good enough and cheap enough that you can now do this at ground level, on consumer-grade cameras, for hyperlocal use cases like a single construction site. Corn: The satellite-to-dirt transition is the thing that makes this newly practical. Herman: Satellite change detection has the luxury of consistent overhead angles, predictable revisit times, and decades of curated datasets. Ground-level construction photos have none of that. The camera gets bumped. Someone parks a truck in front of the thing you're supposed to be monitoring. The models that work in this domain have to be robust to all of that, and the architectures that handle it well are fascinating. You've got two main families — Siamese networks and transformer-based approaches — and the way they handle the correspondence problem between two timestamps is completely different. Corn: Layperson sees trucks. Expert sees process. AI sees anomalies that neither catches. But there's a fourth layer he didn't name, which is the AI also sees a thousand things that aren't anomalies, and someone has to decide which alerts are worth waking up the site manager for. Herman: The triage problem. That's where the deployment engineering matters as much as the model architecture. You can have a technically brilliant change detection system that's completely unusable because it cries wolf forty times a day. Alert fatigue is a real phenomenon — there are studies in medical settings showing that after about twenty false alarms, clinicians start ignoring the system entirely. Construction sites would be even worse because the baseline visual variation is so much higher. Corn: The model has to be good at two things simultaneously: being sensitive enough to catch the subtle stuff an expert would miss, and being disciplined enough not to flag every shadow as an emergency. Herman: Those two goals are in direct tension. Every architectural decision you make is essentially a tradeoff between sensitivity and specificity. The Siamese network approach — which, despite the name, does not involve twins, it's just two identical network branches with shared weights — encodes both images into a joint feature space and then computes a difference map. The idea is that the network learns what "normal" variation looks like and suppresses it, while amplifying deviations that matter. Corn: "Siamese network" is a deeply misleading name for something that has nothing to do with conjoined anything. Herman: It's a historical accident. The original paper used the term because the two branches are identical and joined at the output. Computer vision researchers are not always great at naming things. But the key insight is that by forcing both images through the same feature extractor, you're ensuring that the representation of "a truck" in the Monday photo and "a truck" in the Tuesday photo lands in the same region of feature space, even if the lighting is different or the truck has moved slightly. Then the difference map only highlights things that genuinely changed in a semantically meaningful way. Corn: As opposed to pixel subtraction, which would flag the exact same truck in a slightly different position as a massive change. Herman: Right, because pixel subtraction is spatially rigid. If a truck moves three feet to the left, pixel subtraction says every pixel that was truck and is now dirt is a change, and every pixel that was dirt and is now truck is also a change. You get this ghostly double-image effect where the entire outline of the vehicle lights up. A Siamese network with a well-trained feature space says "that's still a truck, just translated slightly" and suppresses it. Corn: What Daniel's really asking about is a system that can tell the difference between "different truck" and "same truck, different Tuesday. Herman: That's a beautifully concise way to put it. And the transformer-based approaches take this even further. Instead of just encoding both images independently and comparing feature maps, models like ChangeFormer and BIT — the Bitemporal Image Transformer — use cross-attention between the two timestamps. The model literally learns which spatial regions in the Monday image correspond to which spatial regions in the Tuesday image, and it uses that correspondence to decide what constitutes a meaningful difference. Corn: Cross-attention meaning the model is looking at both images simultaneously and drawing lines between them. Herman: In a standard vision transformer, self-attention lets each patch of the image attend to every other patch within the same image. In cross-attention, patches from the Monday image attend to patches from the Tuesday image. So if there's a specific bolt pattern on a piece of equipment that appears in a different location on Tuesday, the attention mechanism can learn to match those patterns across time and recognize it as the same object, just repositioned. That's something pixel subtraction can't do at all, and even Siamese networks can struggle with if the displacement is large. Corn: The attention mechanism is doing a kind of implicit image registration — it's aligning the two photos in feature space without anyone having to manually mark corresponding points. Herman: And it's why these transformer architectures have been pushing benchmark numbers up so dramatically. ChangeFormer, which came out in twenty twenty-two, hit about eighty-five percent intersection-over-union on the LEVIR-CD dataset, a standard benchmark for building change detection from satellite imagery. By twenty twenty-six, fine-tuned variants of DINOv2 are pushing past ninety percent. That's a massive jump in a few years. Corn: For the non-computer-vision people listening, intersection-over-union is basically a measure of how precisely the model's change map matches the ground truth. Eighty-five percent means it's mostly right. Ninety percent means it's right enough that you'd actually trust it in production. Herman: And the DINOv2 part is important for understanding why this became practical for construction sites specifically. DINOv2 is a self-supervised vision transformer from Meta AI trained on a huge corpus of diverse images without needing labels. It learns extremely rich visual features that transfer well to downstream tasks. The shift in twenty twenty-five and twenty twenty-six has been toward using these frozen foundation models as the backbone, then training only a lightweight change detection head on top. Corn: Frozen meaning you don't update the main model weights. You just train the new little bit you've attached. Herman: The practical implication is huge. Instead of needing thousands of labeled change masks — which are expensive and tedious to create, someone has to manually outline every region of change in hundreds of image pairs — you can fine-tune on as few as fifty to a hundred labeled pairs for your specific domain. The foundation model already knows what edges, textures, objects, and spatial relationships look like. The change detection head just needs to learn what kinds of feature differences matter for your particular use case. Corn: For a construction company that wants to monitor three sites, they don't need to hire a team of annotators. They can label a couple weeks' worth of images themselves and get a working system. Herman: That's the promise. And it's why I think Daniel's question is so timely. The barrier to entry has collapsed in the last eighteen months. You can grab a pre-trained DINOv2 model from Hugging Face, add a change detection head — there are reference implementations for both the Siamese and transformer variants — and start experimenting with your own site photos. The TorchGeo library, which has been maturing since twenty twenty-four, bundles a lot of this together with data loaders designed for geospatial and temporal imagery. Corn: Daniel mentioned he's interested in platforms and models, not just the theory. So if someone's listening and wants to actually build this, what's the starting point? Herman: The landscape splits roughly into commercial and open-source. On the commercial side, you've got platforms like SiteAware, which is specifically built for construction progress tracking and safety monitoring. They use proprietary models trained on construction imagery. Orbital Insight does similar work but from the geospatial side — they started with satellite and have been moving toward ground-level integration. These are polished, supported products, but they're black boxes and they're not cheap. Corn: On the open-source side? Herman: TorchGeo is probably the best entry point right now. It's a PyTorch domain library for geospatial data, and it includes pre-trained change detection models, standard datasets like LEVIR-CD+, and utilities for handling the preprocessing pipeline — things like perspective correction and histogram equalization that you absolutely need for ground-level photos. The Hugging Face model hub also has cards for ChangeFormer and BIT that you can download and fine-tune. And there's a growing ecosystem of tutorials and Colab notebooks specifically for temporal change detection. Corn: I want to circle back to something you mentioned earlier about the human filtering mechanism — the idea that experts see what they expect to see. That feels like the core insight that makes this whole thing more than just a neat technical demo. Herman: It's where the real value lives. There's a well-documented phenomenon in radiology — published studies show that radiologists miss twenty to thirty percent of nodules when comparing temporal chest X-rays. These aren't subtle findings in many cases. They're visible. But the human visual system is optimized for change blindness in certain contexts, and when you're looking for specific expected changes, you can completely overlook unexpected ones. Corn: Change blindness is the thing where you don't notice the gorilla walking through the basketball game because you're counting passes. Herman: The invisible gorilla experiment, exactly. And the construction site equivalent would be an engineer reviewing daily photos specifically to verify that the scheduled excavation reached the planned depth, and completely missing that a support beam has developed a crack in the corner of the frame. Their attention was on the depth measurement. The crack was visible but unexpected, so it didn't register. Corn: The AI's lack of expectations becomes a feature, not a bug. It doesn't know what's supposed to be happening, so it doesn't filter out the unexpected. Herman: Which is both the upside and the noise problem we've been circling. The model flags everything that deviates from its learned norm, and most of those deviations are inconsequential. The art is in the thresholding and the triage pipeline. You need a system that scores anomalies by severity, groups related changes so you're not getting fifty alerts for the same event, and presents the top candidates to a human reviewer in a way that makes efficient use of their attention. Corn: Daniel also mentioned medical imaging as a parallel use case. Same architecture, different domain? Herman: Fundamentally the same task — temporal feature comparison — but the noise profiles are completely different. Medical images have controlled lighting, standardized positioning, and consistent acquisition protocols. Construction photos have wild variance in illumination, weather, camera angle, and occlusion. So a model trained on construction data would need significant adaptation for medical use, but the underlying architecture transfers. You'd swap the preprocessing pipeline — lung field segmentation and intensity normalization instead of perspective rectification and shadow removal — and fine-tune on medical image pairs, but the Siamese or transformer backbone stays the same. Corn: That's actually a useful litmus test for whether you understand the problem deeply. If you can articulate exactly what changes between domains and what stays the same, you've probably grasped the architecture. Herman: What stays the same is the core insight: meaningful change detection requires learning a feature space where "same but different conditions" maps close together and "actually different" maps far apart. Everything else — the specific architecture, the preprocessing, the threshold tuning — is engineering around that central idea. Corn: If we're going to unpack this properly, I think we should start with what the model actually looks like under the hood. You've name-dropped Siamese networks and transformers, but I want to understand how the attention mechanism specifically handles that correspondence problem you mentioned — matching the Monday bolt pattern to the Tuesday bolt pattern across different positions and lighting. Herman: Before we get into the attention mechanics, I think it's worth defining the problem precisely, because "visual change detection" means different things in different contexts, and Daniel's construction site scenario is actually a specific and unusually difficult variant of it. At its core, you're comparing two or more photographs of the same scene taken at different times, and you want to identify differences that are semantically meaningful. Not every pixel that changed value. Not every shadow that shifted. You want to know: did something actually happen here that matters? Herman: It's the whole game. What matters on a construction site is completely different from what matters in a medical scan or an ecological survey. On a construction site, a new crack in concrete matters. A different truck parked in the same spot probably doesn't. But the model doesn't know that innately — it has to be taught, either through labeled examples or through the way you structure the task. Corn: Daniel's framing was helpful here because he explicitly named the three viewing layers. The layperson sees trucks and dirt. The expert sees process stages — they know that when the auger rig disappears and a rebar truck shows up, they've moved from drilling to reinforcement. And then the AI layer, which in theory catches things neither human layer noticed. Herman: There's a subtlety in that third layer that's easy to miss. The AI isn't just seeing more detail than the expert. It's seeing differently. The expert's knowledge is both an asset and a filter. They know what to look for, which means they also know what to ignore. The AI has no such filter. It's equally attentive to every square inch of every frame, every day, without fatigue or expectation. That's why it can spot the crack the engineer overlooked. But it's also why it flags the pigeon. Corn: The pigeon is the price of admission. Herman: The pigeon is absolutely the price of admission. And this is why ground-level photos are so much harder than satellite imagery. Satellites pass overhead at predictable times with consistent angles and minimal atmospheric interference. Your daily site photo might be taken at seven AM on a sunny Monday and four PM on a rainy Tuesday. The shadows are completely different. The dirt is darker because it's wet. The camera might have been nudged two degrees to the left by a passing excavator. All of that is noise the model has to learn to ignore. Corn: The fundamental challenge isn't really "spot the difference." It's "spot the difference that isn't explained by lighting, weather, camera angle, or normal daily variation. Herman: That's a much harder problem than it sounds, because "normal daily variation" is itself a learned concept. What's normal on Monday might not be normal on Friday. What's normal during excavation isn't normal during foundation pouring. The model has to build a representation not just of the scene, but of the scene in the context of where the project is in its timeline. Corn: Which means temporal context matters, not just pairwise comparison. Herman: Right, and that's where things get interesting architecturally. A simple pairwise model looks at Monday and Tuesday and flags differences. But a sequence-aware model looks at the whole week and learns that the shadow always moves across the pit between eleven AM and two PM, so a dark patch in that region at one PM isn't an anomaly — it's Tuesday. That kind of temporal consistency modeling is what separates a research demo from something you'd actually deploy. Herman: Let's talk about what's actually happening inside these models, because the architecture choices directly determine what the system can and can't handle. You've got two dominant families. First, Siamese networks — two identical branches with shared weights, each encoding one of the two images into a feature space, then you compute a difference map between those feature representations. The key word is "shared weights." Both images go through exactly the same mathematical transformation. Corn: The network is forced to represent "truck under morning sun" and "truck under afternoon cloud" using the same vocabulary. Herman: That's exactly the right way to think about it. The shared weights create a consistent feature language. A shadow isn't a different object — it's the same object under different illumination, and the network learns to encode it that way. Then the difference map only activates when the semantic content actually changed, not when the lighting did. Corn: The alternative family? Herman: Transformer-based architectures — ChangeFormer, BIT, the Bitemporal Image Transformer. These don't just encode both images independently and subtract. They use cross-attention to explicitly model the relationship between the Monday image and the Tuesday image. Every patch in Monday's photo can attend to every patch in Tuesday's photo, learning which spatial regions correspond across time. Corn: It's building a correspondence map before it even starts looking for changes. Herman: That's the crucial distinction. In a Siamese network, the correspondence is implicit — you hope the shared feature space handles it. In a transformer with cross-attention, the correspondence is explicit. The model literally computes attention weights between Monday-patch-i and Tuesday-patch-j for every pair of patches. If a bolt pattern appears three inches to the left on Tuesday, the attention mechanism learns to connect those two regions and say "same thing, different position." Then the change detection head operates on the aligned representations. Corn: Which means the model can handle much larger displacements without breaking. Herman: A Siamese network with simple feature subtraction starts to fail when objects move more than the receptive field of the convolutional kernels. A transformer with cross-attention has a global receptive field by design — every patch can attend to every other patch. That's why ChangeFormer hit eighty-five percent IoU on LEVIR-CD in twenty twenty-two, and why fine-tuned DINOv2 variants are now above ninety percent. Corn: DINOv2 enters the picture how? Herman: This is the big shift of the last eighteen months. Instead of training a change detection model from scratch — which requires thousands of painstakingly labeled image pairs — you take a pre-trained vision transformer like DINOv2 or SAM, freeze its weights entirely, and attach a lightweight change detection head on top. You only train that head, on maybe fifty to a hundred labeled pairs for your specific domain. The foundation model already knows what edges, textures, objects, and spatial relationships look like. You're just teaching it what kinds of feature differences constitute a meaningful change for your use case. Corn: Fifty pairs is remarkably little data. Herman: It's almost absurdly little, and it's why this has become practical for small construction firms rather than just well-funded research labs. You spend a week labeling images from your own site, fine-tune over a weekend, and you've got a working system. The TorchGeo library bundles pre-trained backbones with change detection heads and data loaders that handle the preprocessing — perspective correction, histogram equalization, all the stuff that makes ground-level photos usable. Corn: Let's go back to the noise problem, because I think that's where most real-world deployments actually fail. You mentioned earlier that even feature-difference approaches can flag shadows and moving leaves as changes. Herman: This is where the naive approaches completely fall apart. Imagine our construction site on a partly cloudy day. Between the Monday photo and the Tuesday photo, a cloud moves across the sun, and a shadow sweeps across the excavation pit. A pixel-difference model lights up that entire shadow region as change — hundreds of square meters of false positive. A basic Siamese network with feature subtraction does better, but can still get confused if the shadow edge aligns with a real structural feature. Corn: Because the feature representation of "pit wall in shadow" might differ from "pit wall in sunlight" in ways the network hasn't learned to normalize. Herman: The solution is temporal consistency modeling — using not just two frames but a sequence. If you have Monday through Friday, you can use three-dimensional convolutions that operate across the time dimension, or recurrent connections that maintain a running representation of what's normal at each spatial location. The model learns that this particular corner of the pit goes dark every day at two PM, so darkness at two PM on Thursday isn't an anomaly. Corn: It's learning the shadow's schedule. Herman: That's a wonderfully simple way to put it. And once the model has that temporal baseline, it can distinguish between "the shadow that always moves through here" and "the crack that appeared here yesterday and hasn't gone away." The shadow is periodic and predictable. The crack is persistent and novel. Those are fundamentally different signals, but you can only separate them if you're looking at more than two frames. Corn: The pairwise approach has an inherent ceiling on precision that sequence modeling breaks through. Herman: Right, and in practice the best deployed systems use a hybrid. They maintain a temporal baseline model that learns normal variation patterns across days or weeks, and then a pairwise change detector that compares today's image against that baseline rather than against yesterday's specific image. The baseline absorbs the shadows, the weather, the normal vehicle traffic. The pairwise comparison only fires on deviations from normal. Corn: Here's the thing that I think Daniel was really driving at — even with a perfectly tuned model that handles shadows and pigeons and weather, you've still got this deeper question about human expertise. The model catches something the engineer missed. What does that actually look like in practice? Herman: There was a fascinating pilot on a highway construction project last year — a team used a fine-tuned DINOv2 model monitoring overnight site photos. One morning the model flagged a crane that had been repositioned about forty feet from where it was supposed to be. The site manager hadn't authorized the move, and when the human reviewers looked back at the images, they'd all skimmed right past it. They assumed it was part of the day's planned work. Corn: The AI's ignorance of the plan was the feature. It didn't know the crane was supposed to be there, so it didn't mentally file it under "expected. Herman: That's the expert blind spot in action. The engineer's mental model of the project timeline actually prevented them from seeing the deviation. There's a whole literature on this in radiology — published studies consistently show radiologists miss twenty to thirty percent of nodules in temporal chest X-ray comparisons. These aren't ambiguous edge cases. They're visible findings that the expert's attention simply didn't land on because they were focused elsewhere. Corn: Which means the AI isn't replacing the expert — it's functioning as a second pair of eyes that doesn't share the expert's assumptions. Herman: That's the correct framing. The best deployed systems are triage layers, not autonomous decision-makers. The pipeline in practice looks like this: fixed camera with GPS and timestamp captures the daily photo, automated upload kicks off preprocessing — perspective correction, histogram equalization, all the normalization we talked about — then the model runs inference and produces an anomaly score for each detected change region. Those scores feed into a review queue prioritized by severity. Corn: Someone has to decide where to set the threshold for what reaches a human. Herman: That's the single most important design decision in the whole system. Set it too low and you flood the engineer with false positives — shadow alerts, pigeon alerts, truck-parked-slightly-differently alerts. Alert fatigue sets in within days. Set it too high and you miss the crack propagating in the corner of the frame. For construction safety applications, most teams err on the side of over-alerting because the cost of a missed structural failure is catastrophic. For progress tracking, you tune for precision because the cost of a false alarm is just wasted attention. Corn: The threshold isn't a technical parameter — it's a business decision dressed in technical clothing. Herman: It varies by domain. Medical imaging thresholds are calibrated completely differently because the cost structure is different. A false positive in a chest X-ray means an unnecessary follow-up scan. A false negative means a missed cancer. Those have wildly different weights than a construction site where a false positive means someone looks at a photo for thirty seconds and dismisses it. Corn: Daniel asked about platforms too. If someone wants to actually build this rather than just admire the architecture from a distance, what's the landscape? Herman: It splits into commercial and open-source. On the commercial side, SiteAware is the dominant player for construction-specific monitoring — they've built proprietary models trained on years of construction imagery, and they handle the full pipeline from camera deployment to alert dashboards. Orbital Insight does similar work but from the geospatial angle, and they've been expanding into ground-level integration. These are polished products, but they're black boxes and priced accordingly. Corn: For someone who wants to get their hands dirty? Herman: TorchGeo is the best entry point right now. It's a PyTorch domain library that's been maturing rapidly since twenty twenty-four, and it bundles pre-trained change detection models with geospatial data loaders and preprocessing utilities. The LEVIR-CD+ dataset is available for fine-tuning on building and construction scenes specifically. And the Hugging Face model hub has ChangeFormer and BIT model cards you can download and adapt. You can go from zero to a working prototype in a weekend if you have labeled image pairs. Corn: Which brings us back to the cross-domain question Daniel raised. Construction site to medical imaging — same architecture, different everything else? Herman: The underlying task is identical — temporal feature comparison — which is why the Siamese and transformer backbones transfer so cleanly. But the preprocessing pipelines are completely different. Medical imaging needs lung field segmentation and intensity normalization because you're dealing with calibrated X-ray machines in controlled environments. Construction needs perspective rectification and shadow removal because you're dealing with outdoor chaos. The noise profiles are almost inverted. Corn: Controlled lighting versus "the sun does what the sun wants. Herman: That means a model fine-tuned on construction data won't just work out of the box on chest X-rays — you need to swap the preprocessing and retrain the change detection head. But the DINOv2 backbone, the attention mechanism, the fundamental approach of learning what constitutes a meaningful difference in feature space — all of that transfers. It's the same engine with a different body. Corn: Which is a testament to how general the problem actually is. The specifics change but the structure doesn't. Herman: That's what makes this moment so interesting. The foundation model revolution of the last two years means you're no longer building bespoke systems for each domain from scratch. You're adapting a general capability. The construction firm and the radiology lab are solving the same underlying problem with different window dressing. Corn: If someone listening wants to actually try this — not just nod along with the architecture discussion — what's the concrete starting point? Herman: Grab a frozen foundation model. DINOv2 or SAM. Don't touch the backbone weights. Attach a lightweight change detection head — there are reference implementations for both Siamese and transformer variants on Hugging Face — and fine-tune on your own domain data. You need as few as fifty to a hundred labeled image pairs. That's a weekend of annotation for most construction sites. Corn: Fifty pairs feels almost suspiciously low. Herman: It's the foundation model carrying the load. DINOv2 already understands edges, textures, object boundaries, spatial relationships. You're not teaching it to see. You're teaching it what kinds of differences matter for your specific site. The head is tiny compared to the backbone — training converges fast and doesn't overfit the way a from-scratch model would on that little data. Corn: The second thing you said earlier about thresholds being a business decision — that feels like the thing everyone gets wrong on their first deployment. Herman: Because they treat it as a technical parameter to optimize, not a risk decision to make. You have to sit down and ask: what's the cost of missing something versus the cost of crying wolf? On a construction site where a missed crack could mean a structural failure, you set the threshold low. You accept more false positives because the alternative is catastrophic. For progress tracking — where you're just verifying that Tuesday's excavation reached the planned depth — you tune for precision. Alert fatigue is real and it kills adoption faster than any model bug ever could. Corn: The same model, same site, might need two different thresholds depending on what question you're asking. Herman: Safety monitoring and progress tracking are different tasks even though they run on the same imagery. Treating them as one pipeline with one threshold is a recipe for either missed risks or ignored alerts. Corn: For the developer who wants to experiment before committing to a deployment? Herman: TorchGeo is the cleanest on-ramp. It bundles pre-trained change detection models with data loaders that handle the preprocessing — perspective correction, histogram equalization, all the normalization steps that make ground-level photos usable. The Hugging Face hub has ChangeFormer and BIT model cards ready to download. Start with a simple Siamese ResNet baseline so you understand the mechanics, then swap in a vision transformer backbone and watch the improvement. The jump is not subtle — you'll see it in the IoU numbers immediately. Corn: Siamese ResNet as training wheels, then DINOv2 as the real thing. Herman: The training wheels version still beats pixel subtraction by a mile. It's not that the simple approach is bad. It's that the foundation model approach is so much better that it changes what's practical to build. Corn: The question that sits in my head after all of this is what happens when these systems get good enough and cheap enough that they stop being optional. Not "should we deploy this" but "can we legally not deploy this. Herman: The liability question. It's the one that keeps lawyers up at night, and I think it's going to hit construction before it hits medicine, ironically. Medical AI has the FDA as a gatekeeper. Construction doesn't have an equivalent regulatory bottleneck. If a model exists that demonstrably catches structural risks that human inspectors miss, and a firm chooses not to use it, and something fails — is that negligence? Corn: "The algorithm saw it and you could have too, if you'd bothered to turn it on. Herman: We already have the precedent in other industries. If you're a radiologist and a CAD system flags a nodule you dismissed, that's discoverable in a malpractice case. The standard of care shifts as the technology becomes widely available. For construction, I think we're maybe three to five years from insurers starting to ask whether change detection was part of the monitoring regime. Corn: The flip side — when the model misses something, who owns that? The contractor who deployed it? The platform vendor? The open-source developer whose model weights they downloaded? Herman: That's the part nobody has a good answer for yet. With commercial platforms like SiteAware, there's presumably a contract and an SLA. With an open-source model fine-tuned by the contractor's summer intern — that's a much murkier chain of responsibility. Corn: Which makes me think the real next frontier isn't just better models. It's real-time detection from drone video feeds. You're not comparing two photos anymore. You're processing a continuous stream. Herman: That's a fundamentally different problem. Discrete daily pairs let you batch process, run heavy models, take your time. Continuous video means you need models that can run inference at thirty frames per second on edge hardware, maintaining a running representation of what's normal and flagging anomalies as they happen. The latency budget goes from "sometime tonight" to "before the excavator finishes its swing. Corn: The payoff is you catch things in time to do something about them, not in time to write a report about what already went wrong. Herman: That's the dream. And it's going to require a whole different class of architectures — lightweight temporal models that can run on a drone's onboard compute, probably some variant of recurrent vision transformers with aggressive quantization. The research is happening now, but we're not there yet. Corn: Something to watch. And on that note — Hilbert, you've got the floor. Herman: Now: Hilbert's daily fun fact. Hilbert: In nineteen oh eight, astronomers at the Atacama Desert's newly built Mills Observatory attributed a peculiar lunar wobble to gravitational interference from an undiscovered asteroid they named "Hidalgo's Shadow." Forty years later, a graduate student proved the effect was actually longitudinal libration — the Moon's own orbital geometry, not an external body. The asteroid never existed. Corn: Forty years of phantom asteroid. That's commitment to being wrong. Herman: This has been My Weird Prompts. If you enjoyed the episode, leave us a review wherever you listen — it helps more people find the show. I'm Herman Poppleberry. Corn: I'm Corn. We'll be back next week.