/ 01↗
Detection
Axis-aligned bounding boxes around individual leaves at canopy scale, robust to occlusion and motion blur.
0.943 mAP@50
live · model onlinebuild 3.1.04
amravati · maharashtra—
◉ rudra/xb3·1arch · transformer-yolo hybridparams · 11.4Mmap@50 · 0.943latency · 27ms / frame↳ scroll for capabilities
/ 01 · precision agriculture
See every leaf.
Diagnose the field
before it whispers.
Rudra XB3·1 is an edge-deployable vision model that detects, segments, and classifies leaves in real time — flagging early disease, measuring canopy stress, and counting yield before the human eye registers a thing.
- 27ms
- Inference
- 36 fps
- Throughput
- ≤ 4W
- Edge target
rec · cam_03 · 1080p
fps 36 · 27ms
drop /public/leaf-detection.mp4 · 15s · loop
leaf · healthy0.97
leaf · early blight0.88
leaf · healthy0.93
3 detections · 1 anomaly · field 04-A
n 20.93° · e 77.75°
Sample output — Rudra XB3·1 inferring on a Jetson Orin Nano. Bounding boxes generated client-side; the underlying clip shows raw canopy footage from a Vidarbha cotton trial.
model
xb3·1-prod
◇ trained on 2.4M field-captured frames◇ deployed across 14 districts◇ vidarbha cotton trials · 94.3% accuracy◇ tea estates · darjeeling · live◇ iso 27001 · soc 2 type ii◇ runs on jetson · coral · raspberry pi 5◇ open weights for academic research
/ 02 · capabilities
Six tasks.
One forward pass.
XB3·1 is a multi-head architecture. Detection, segmentation, disease classification, growth-stage estimation, canopy density, and oriented count — all emitted from a single inference call. No model swapping, no second-pass overhead, no cloud round-trip.
/ 02↗
Segmentation
Pixel-precise leaf masks for canopy area calculation and per-leaf disease scoring.
0.91 IoU
/ 03↗
Disease classification
12 disease classes across cotton, soy, tea, grape, and tomato — plus a 'novel' fallback for unknowns.
12 classes
/ 04↗
Growth stage
Estimates BBCH stage from foliar morphology — crucial for irrigation and spray scheduling.
BBCH 11–89
/ 05↗
Canopy density
Per-frame leaf area index with a confidence band, suitable for stress-mapping over time.
LAI ± 0.07
/ 06↗
Oriented count
Rotated-bbox counting for trellised crops where standard detection over-merges adjacent leaves.
± 2.1% error
/ 03 · pipeline
Frame in. Decision out.
Twenty-seven milliseconds in between.
- stage · 010 ms
Capture
RGB or RGB-NIR feed from a fixed-pole or drone-mounted camera. 1080p at 36 fps with adaptive auto-exposure.
- stage · 024 ms
Preprocess
Letterbox to 640×640, gamma-correct under harsh sun, and stabilise against wind sway via inter-frame optical flow.
- stage · 0321 ms
Inference
Multi-head XB3·1 backbone runs detection, segmentation, and classification heads in a single forward pass.
- stage · 042 ms
Decide
Threshold against field-specific priors. Trigger SMS, irrigation valves, dashboard pin, or all three.
deployment · python
from rudra import XB3
model = XB3.load("xb3-1.prod") # 11.4M params, ~46 MB
model.to("edge:jetson-orin-nano") # or "coral", "rpi5", "cpu"
for frame in field_camera.stream():
out = model.infer(frame) # 27 ms median
if out.has_anomaly():
alert(out.geo, out.disease) # → SMS / dashboard / pump valveruns on
- ▸NVIDIA Jetson Orin Nano
- ▸Google Coral TPU
- ▸Raspberry Pi 5 (4GB)
- ▸x86 CPU (AVX2 fallback)
- ▸Browser (WASM, experimental)
/ 04 · field notes
Logged from the
soil up.
We don’t train in a sterile dataset and pray. XB3·1 is shaped by partnerships with farmer collectives, agronomy researchers, and the kind of edge cases you only find at sunrise on day 47 of a monsoon.
log · vidarbha-04afeb 2026
“Picked up early blight on three plants in row 12 — eight days before any of us would have spotted it walking the field. That window saved the crop.”
log · darjeeling-teajan 2026
“We hung two cameras over a single garden and the model tracked flush stages cleaner than our pluckers could log them. Full stop.”
log · nashik-grapemar 2026
“Latency on the Jetson stayed under 30ms even at 38°C ambient. That’s the number that matters when you’re running pumps off the same inference call.”
/ 05 · request access
Put a model
in the field.
We’re onboarding a small cohort of farms, cooperatives, and research labs for the XB3·1 production release. Drop your email and we’ll send a deployment kit within seven days.