Why “lightweight” computer vision matters
Most teams don’t need a research lab or a full ML Ops organization to ship computer vision. What they need is a repeatable way to train a model, deploy it reliably, and keep data secure—especially when the inputs are sensitive (game film, contracts, claims, financial docs, facility footage).
Our University of Miami (UM) + Lid Vizion collaboration with the REMADE Institute proved a simple point: when your model is efficient enough, you can put vision into real workflows without expensive infrastructure—and you can reuse the same blueprint across industries.
The project: identifying materials at the source (UM + Lid Vizion + REMADE)
Recycling programs suffer from contamination largely because people are unsure what belongs where. Instead of relying on broad awareness campaigns, the project built an API that recognizes an item in an image and returns a clear instruction based on local program rules and the consumer’s location.
The system was designed as a modular service that can power a web app or smartphone app: scan → identify → verify against rules → respond.
Architecture: a practical CV API that scales without a big team
We intentionally used an architecture that a small technical team can ship and operate:
React front-end for fast iteration and a simple user workflow
API Gateway for API management, authentication, throttling, and usage plans
AWS Lambda to run inference on demand (serverless, scalable, cost-efficient)
Amazon S3 for model artifacts and storage
MongoDB Atlas as the rules + response layer (accepted materials, program definitions, logic)
This “model + rules engine” pattern is the key: the model detects what it sees, and the database determines what it means for the user’s specific context.
Dataset and training: what actually makes the model work
Models don’t fail because “AI is hard.” They fail because the dataset doesn’t match reality. For this case study, we created the UM‑LV Recycling dataset focused on high-value plastics and common consumer electronics that frequently contaminate streams:
Plastics: PET clear (#1) and HDPE translucent (#2)
Electronics: batteries, remotes, gamepads, calculators
The dataset included ~2,842 labeled images in the research report, captured across different devices and environments. To improve robustness, the project summary also highlights common augmentation techniques (rotation, flipping, adding noise, cropping), expanding the effective training set to >10,000 images.
Model bake-off: performance vs. deployability
We tested multiple state-of-the-art models and compared them using standard metrics (precision/recall/F1 and mean average precision at IOU 0.5). Here’s what mattered most for deployment:
Accuracy must be high enough to avoid costly misclassification (especially for hazards like batteries)
Latency must be low enough for a real-time user experience
Model size + efficiency must support serverless/edge paths (not just GPU servers)
Results snapshot (mAP @ 0.5 IOU)
Faster R‑CNN: 83.21%
ResNet‑50 (single-shot detector): 92.36%
MobileNetV2: 96.15%
EfficientDet: 96.35% (selected)
Two models stood out: MobileNetV2 and EfficientDet. EfficientDet slightly led overall and became the selected model for the API, while MobileNetV2 validated the “lean model” approach that’s well-suited to edge devices.
What “edge-ready” really means
Edge doesn’t have to mean “no cloud.” Edge-ready means your model and system are efficient enough to run close to the data when it matters:
On-device inference for privacy (confidential documents, proprietary video)
On-prem or facility inference to avoid transferring large video streams
Serverless inference to scale without constant infrastructure work
This is why lightweight models matter. They preserve your options: start with an API for speed, then move inference closer to the edge as requirements tighten.
Where this goes next: sports game film (baseball, basketball, football, soccer)
Sports is a perfect example of why edge-friendly vision wins. Game film is:
Large (video hours add up fast)
Time-sensitive (coaches need answers quickly)
Confidential (teams don’t want footage leaking)
The same blueprint applies—only the objects and rules change. Lightweight models can power:
Object detection + tracking: ball, players, officials, key zones
Event detection: shots, passes, tackles, possessions, transitions
Spatial + tactical overlays: spacing, lanes, pressure, defensive shape
Workflow automation: clip creation, tagging, and searchable film libraries
For teams without ML Ops, the practical path is: start with a narrow set of detections, instrument feedback loops from coaches, and expand category coverage as ROI becomes obvious.
Confidential docs: “review like a human” without sending data everywhere
Now apply the same edge-ready approach to documents. In regulated and high-trust workflows (real estate, logistics, finance, and more), companies need models that can operate under strict privacy and compliance constraints.
Lightweight vision + document intelligence can support:
Classification: identify doc types (leases, invoices, bills of lading, claims, statements)
Extraction: pull key fields and normalize them
Verification: detect inconsistencies vs. policy or prior records
Audit trails: log what the model saw, what it suggested, and what a reviewer approved
When teams treat “document review” as a repeatable pipeline (not a one-off script), it becomes safer, faster, and more defensible for audits.
Impact (and why it’s bigger than one use case)
The REMADE project summary estimates meaningful carbon savings from reducing contamination and improving participation—up to 22,765 metric tons of CO₂ annually for Miami‑Dade County, and ~4.5 million metric tons nationally if participation improves by 15%.
But the real takeaway is the technical playbook: efficient models + a clean API + a rules layer = deployable computer vision for organizations that don’t want to become an ML infrastructure company.
Want to apply this to your use case?
If you’re working with game film, facility video, or confidential documents—and you want a lightweight, secure path to production—reach out. We’ll help you map:
what to detect/extract first (highest ROI)
which model family fits your constraints (latency, privacy, cost)
how to deploy without building a big ML Ops function
Let’s talk. Bring a sample of your footage/documents and your rules—our job is to turn that into a deployable edge-ready workflow.
References
University of Miami IDSC: AI recycling project at the source
REMADE Institute project summary: 20-01-RR-4042