Professional kitchens throw away food worth a large share of their food purchases, but rarely know what, when or why. A smart bin combines a scale with a camera: staff drop waste in, the system identifies the item, logs weight, cost and CO2e and ranks the causes. Dashboards turn this into concrete changes (smaller batches, menu tweaks, portion sizes) and into auditable data for sustainability reporting.
Problem & value
- Today: waste is estimated, weighed manually on paper, or only measured in occasional audits. Chefs know total bin volume, not which dish, prep step or day causes it. Manual logging fades after a few weeks.
- AI value: the camera removes manual data entry (the scale gives weight, the vision model gives the item and category), so measurement continues without extra work. Visibility then drives behaviour change; vendors report substantial reductions in food waste for their customers, but these are vendor-published case figures and should be validated in a pilot with your own baseline [3][4].
- Regulatory driver: the revised EU Waste Framework Directive (Directive (EU) 2025/1892) introduces binding food-waste reduction targets for 2030 (processing/manufacturing and, per capita, retail, restaurants, food services and households), with Member States to measure and report food waste on a common methodology. This pushes caterers and hospitality groups towards quantified data [1][2].
- Cost logic: every kilogram avoided saves purchase cost, labour and disposal fees, so payback is mostly driven by food cost share and volume.
Variants
- Production kitchens (canteens, hospitals, care homes, schools, airline/contract caterers): spoilage, overproduction and plate waste by meal period.
- Hotels and buffets: buffet leftovers vs. prep waste; guest-facing signage and portion adaption.
- Restaurants and chains: per-site benchmarking against peers, central procurement.
- Retail and supermarkets (bakery, deli, fresh counters): shrink tracking at store level, link to markdown tools.
- Plate-waste tracking in canteens: tray-return camera (no scale) to estimate leftovers per dish.
- Duplicate/merged candidates: retail-ecommerce smart-bin measurement and the personal-life listing (adds Kitro, CH) are covered here.
- Private use: household food-waste logging apps exist but are not smart-bin based.
How it works
- waste dropped into bin on scale
- scale triggers capture (weight stable)
- camera image(s)
- edge classifier / detector: food item, category (protein, carb, veg, bakery, ...)
- staff tags reason/stage on tablet (spoilage, overproduction, trim, plate) - optional, often a few taps
- cloud: weight × item cost × emission factor = cost and CO2e
- dashboards, daily/weekly reports, anomaly alerts
- recommendations: production quantities, menu changes, purchasing
- ESG / waste-reporting export
- Recognition: image classification and detection models trained on mixed, partly eaten and occluded food; the model proposes a label, staff confirm or correct. Corrections feed retraining. Newer systems use VLMs for long-tail items and free-text reasoning.
- Weight and cost: the scale is the ground truth for mass; cost uses item-level prices from invoices/POS, CO2e uses emission factors per food category.
- Human-in-the-loop: chefs/managers decide on actions; staff correct labels. The AI recommends, it does not change production automatically.
- Analytics: forecasting of covers/portions (see PRED-001 family) to compare planned vs. actual production.
Data preparation
The system creates most of its own data, but it needs reference data to turn kilograms into cost and CO2e: food cost per item (supplier invoices as PDF or paper, kitchen software, ERP or POS), menus and production plans (Excel, kitchen software, printed sheets) and covers per meal period (POS or booking system). Typical gaps: item names differ between purchasing, POS and recipes; prices exist per supplier unit, not per kg; plans live on paper or in the chef's head; there are no agreed waste reasons (prep, spoilage, overproduction, plate).
Preparation steps: - Map the 50-100 most important items to purchasing articles and a cost per kg in one spreadsheet. - Fix the reason categories before go-live (PREP-019); the same words for every shift. - Get invoices as e-invoices or scan them (PREP-006); Paperless-ngx or the vendor's import can read them. - Put production plans into one tidy table (PREP-011). - Use the first weeks of staff corrections as a labelled set (PREP-020, PREP-021) to improve recognition. - Frame the camera on the bin only, so that no staff are visible.
Effort: low, roughly 15-20 percent of the project. The larger part is hardware rollout and building the habit of tapping reasons. For a structured check see GUIDE-003.
Implementation paths
Which path for whom
- smb: A - integrated, proven smart-bin products with scale, camera and dashboards exist and the process is standard for any kitchen.
- enterprise: A - food-waste tracking is a commodity function; caterer and hotel chains profit from vendor benchmarks and rollout support.
- public: B - canteens, hospitals and municipal kitchens often need tendered, EU-hosted solutions; a generic scale and camera with an EU-region API keeps procurement open and avoids lock-in.
Path A - Buy (recommended)
- Winnow (UK, Winnow Vision with camera and scale; operates in Europe), Orbisk (NL, EU), Kitro (CH, also active in DACH), Leanpath (US, with European presence; scale plus camera tracking). Further tools exist for waste logging without cameras (e.g., Too Good To Go business tools for surplus sale; not tracking).
- Typical offering: hardware (scale, camera, tablet) on subscription, onboarding, dashboards, consultant support.
- Check: EU data hosting, camera field of view (bin only), integration with POS/ERP, ESG export formats, contract term, support in local language.
- Choose A for almost every kitchen: hardware and models are integrated and proven.
Path B - Build on cloud APIs
- Capture with a generic scale + camera + tablet; send images to a multimodal API (Claude, Gemini, GPT, Mistral Pixtral/Medium class) in an EU region for item labelling with structured output; store weights and labels in a cloud DB (e.g., EU region Postgres) and visualise in Metabase/Power BI.
- Fits sustainability teams with in-house dev capacity or research projects. Latency and cost per image must be managed; cache repeated items.
Path C - Self-host (open source)
- Edge device (Jetson Orin Nano, Raspberry Pi 5 + Hailo, or a mini PC) runs an open detector/classifier: Ultralytics YOLO (check AGPL/enterprise licence), RT-DETR, DINOv2 or SigLIP embeddings with a small head, or an open VLM (Qwen-VL, SmolVLM) for rare items.
- Public data: Food-101 and UEC-Food256/FoodSeg103 for pre-training; kitchen-specific images are needed for good accuracy.
- Stack: MQTT from scale, Postgres/TimescaleDB, Grafana. Hardware: HW1 at the edge; HW2 only if training centrally.
Path D - Custom training
Worth it for chains and caterers with a recurring menu: fine-tune on a few thousand labelled bin images from your own kitchens, using staff corrections as labels. Class sets must include cooked mixed dishes, sauces and packaging contamination. Low-light and steam robustness matter.
State of the art (2026)
- Mixed waste is the hard part: piled, wet, partly eaten and sauce-covered food is visually ambiguous. Systems therefore rely on weight as ground truth plus coarse category classification, not on perfect dish recognition; accuracy claims differ between vendors and should be tested on your own bins.
- Foundation-model backbones (DINOv2/SigLIP-type embeddings, VLMs) reduce the labelled data needed for new menu items; few-shot labelling by staff is now common.
- Beyond the bin: tray-return cameras for plate waste, inventory/spoilage cameras in cold rooms, and links to forecasting and procurement.
- Carbon accounting: CO2e is derived from category-level emission factors, an approximation; treat it as indicative unless the method is documented.
- Limits: staff compliance (waste bypassing the bin), misclassification of similar-looking items, lighting and steam, double counting between stages, and weak causality: dashboards show where waste arises, but menu and process changes still require chefs.
Shared resources & combinations
- Same stack: VIS-006 (food grading, same edge classifier approach), VIS-020 (waste sorting with similar bin-top cameras), VIS-026 (volume and counting).
- Downstream: DATA-012 (ESG/CSRD reporting), PRED-001 (demand forecasting) for production planning.
- Privacy tooling: VIS-030 (blurring) if camera views can include staff.
- Edge hardware (Jetson/Hailo) and fleet management can be shared with other camera use cases in the same building.
Effort, cost & KPIs
- Path A: weeks; hardware and subscription per kitchen; training of staff in the first month is the main effort.
- Path B: 2-4 months for a pilot with dashboards; API cost per image is small compared with the labour it saves.
- Path C/D: 4-9 months incl. data collection; needs ML and edge operations skills.
- KPIs: kg waste per cover or per 100 meals, waste cost as % of food purchases, CO2e avoided, share of waste by cause (prep, spoilage, plate), compliance rate (items logged vs. estimated), payback time.
People & change
Start with management (budget, sustainability goal) and the head chef or kitchen manager as process owner; both define the target, e.g. less waste, not control of staff. Involve IT for network and tablet, and the data owners for prices and plans. End users (cooks, dishwashers) must be involved early: show how the bin works, that it measures food, not people, and that tapping a reason takes seconds. Pilot in one kitchen or one meal period for 4-6 weeks (GUIDE-009).
Works council: depends. The camera looks only at the bin, but the system logs who drops waste when if staff log in. If a works council exists, inform it in good time (§90(1) no. 3 BetrVG) and agree on a rule that reports are per kitchen and shift, not per person; §87(1) no. 6 BetrVG applies as soon as the system is objectively suited to monitor behaviour (GUIDE-004). Public canteens: staff council.
DPO: low risk when no persons are visible; check camera field of view and any user log-ins (GUIDE-005). AI literacy: a short briefing on what the model can get wrong (mixed dishes) supports staff competence under AI Act Art. 4 (GUIDE-006). Typical concerns: "I am being watched", "extra work"; answer with transparency, aggregated results and visible savings shared with the team (GUIDE-008). Humans decide on menu and production changes.
Risks, regulation & compliance
- EU AI Act: minimal risk. Not an Annex III use case; Art. 4 AI literacy applies. Avoid any use that evaluates individual employees' behaviour by biometric inference; the system monitors food, not people (Art. 5 workplace emotion recognition is not involved).
- GDPR: images should capture the bin only. Cameras that record staff faces or hands create personal data and employee monitoring concerns; works-council co-determination applies in Germany (BetrVG §87). Mitigate with narrow field of view, no audio, short retention, blurring or on-device processing, and a DPIA when in doubt.
- Waste reporting: align methodology with the Commission's common food-waste measurement methodology (Delegated Decision (EU) 2019/1597) for comparability [2].
- Other risks: vendor lock-in on data, overclaimed savings, greenwashing in CSRD statements; document assumptions and keep raw weights auditable. Food safety documentation (HACCP) is not replaced by this tool.
Getting started
- Measure the baseline for 2-4 weeks (weigh waste manually per stage) and compute waste cost per cover.
- Pick 1-2 high-waste kitchens for a 3-month pilot with a vendor from Path A.
- Define the camera field of view with staff representatives/works council; document the DPIA.
- Set weekly review meetings where the chef acts on the top three waste items.
- Export results for sustainability reporting and decide on rollout after the pilot.
Sources
- Directive (EU) 2025/1892 amending Directive 2008/98/EC (Waste Framework Directive), EUR-Lex - https://eur-lex.europa.eu/eli/dir/2025/1892/oj
- European Commission - Food waste reduction targets and measurement - https://food.ec.europa.eu/safety/food-waste_en
- Winnow - Winnow Vision for commercial kitchens - https://www.winnowsolutions.com
- Orbisk - AI food waste measurement - https://www.orbisk.com
- Leanpath - Food waste prevention technology - https://www.leanpath.com
- Kitro - Food waste tracking - https://www.kitro.ch
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj