In a peloton shot, 50-100 riders are visible, bibs are on their backs (hidden from you when they're facing away), and they overlap completely. Reading even 10 bibs from a single frame requires superhuman eyes and time. Traditional tagging captures 1-2 bibs and misses the rest.
For race photographers covering Grand Tours (Tour de France, Giro, Vuelta), missing riders means missing story. For event photographers covering sportives and gran fondos, riders who can't find their photo don't buy. A peloton photo with 50 riders tagged to only 2 riders is a missed opportunity for 48 additional sales or editorial value.
Multi-rider bib detection in cycling is the ability to identify and read bib numbers on multiple cyclists simultaneously, with the added complexity that bibs are positioned on the back of the jersey (behind the saddle bag or frame pack) and riders are tightly packed in 3D space with multiple overlap layers.
Professional cycling media relies on identifying the peloton composition from photos. Knowing which teams and riders were in the break, who was dropped, which teams controlled the pace — all requires reading multiple bibs from single frames. For event photographers, the revenue model depends on each rider finding themselves in the crowd.
In specifically:
Cycling bib placement is unique among race sports: bibs are on the back of the jersey, not the front, making them invisible unless you're shooting from behind; peloton density (50-100 riders in an area 10m wide × 30m long) creates multiple overlap layers; riders lean in and out, bibs fold against saddle bags, and combined with frame numbers and helmet numbers, there are multiple ID sources competing for visual space; long-lens shooting from motorcycles compresses perspective and makes bib sizes inconsistent.
Wide peloton shot with 60+ riders visible, shot from motorcycle perpendicular to the road
very common✗ Back bibs are clearly visible on most riders, but they're at varying distances (closest riders 5m away, back of peloton 30m away) and angles (some facing camera, some at 45°, some showing only side). Reading bibs across this range with traditional methods requires either multiple manual passes or AI.
Tight pace-line photo (5-8 riders in tight formation) where back bibs overlap and riders' bodies are nearly touching
very common✗ Even though bibs are visible, they overlap and blend into the mass of road gear. Individual bib separation by eye is nearly impossible. OCR can't separate overlapping text. AI has to reconstruct individual bibs from partial pixel data.
Breakaway shot with 4-5 riders at distance, shot from front at an angle, bibs partially hidden by saddle bags and bottles
common✗ Back bibs are visible but partially obscured by hydration packs and frame bags (which many cyclists now use for storage). Only 2-3 digits of a 4-5 digit bib might be visible. Manual reading is slow; OCR fails; AI has to infer from partial data.
Crowd crossing the finish line with spectators in the frame, 3-4 riders bunched together with spectators filling gaps
occasional✗ Background clutter (spectators, barriers, team staff) creates visual noise that occludes parts of back bibs. Bibs are still visible but the background isn't clean. Traditional detection systems trained on clean photos struggle with cluttered finish line scenarios.
Manual identification from field notes (ask team staff or use known rider appearances)
⚠ Doesn't scale. Grand Tour photographers cover 3000+ images per stage. Manual identification of peloton composition is impossible. Only works for elite professional cyclists where appearances and team colors are known.
Single-rider focus tagging (tag the closest/clearest bib, skip the rest)
⚠ Reduces a peloton of 50 riders to a single photo tag. Complete loss of secondary identification. Sportive photographers lose 49 of 50 potential sales.
Timing system cross-reference (GPS, chip timing to match riders to approximate position/time)
⚠ Only works if you have real-time GPS or chip timing. Most amateur sportives don't have per-rider tracking. Professional teams have positioning data but it's not publicly available.
AI vision detects all human subjects (riders) in the frame, locates the back bib region on each detected cyclist, and reads the bib number independently for each rider. The system understands rider body positioning (forward lean, saddle position) and infers bib location even when partially hidden by bags or compressed by overlap. For back bibs at distance, the AI uses size inference and perspective understanding to maintain reading accuracy.
Key advantage
Simultaneous multi-rider detection from a single frame without manual cross-referencing. A 60-rider peloton photo produces 60 individual bib detections in one pass. The photo gets tagged to every identified rider, enabling both editorial (which teams in the break) and commercial (riders find themselves in the photo) uses.
94-97% per-bib accuracy on tight-packed pelotons with 20-50 riders, moderate distances (5-20m)
Good conditions
87-92% when peloton is spread (50m+ depth), back bibs at extreme foreshorten angles, or saddle bags cover bib regions
Challenging
76-84% with confidence flags when multiple overlap layers, extreme distance variation (5m to 50m), or extreme angle distortion
Worst case
Process your stage negatives. RaceTagger detects every visible bib in each frame and outputs a JSON per-photo: [{photo_id: 'A001.NEF', bibs: [142, 67, 89, 156, ...]}, ...]. Each bib tags that photo to that rider. Output XMP sidecars compatible with Lightroom and Photo Mechanic. Low-confidence bibs are flagged for quick visual verification — typically 8-12% of detected bibs.
| Metric | Manual | Basic OCR | AI Vision (RaceTagger) |
|---|---|---|---|
| Processing time (3000 peloton photos per stage, avg 40 riders per photo) | 20-40 hours (cross-reference timing data, multiple passes, team info) | 60-90 minutes (but 1-2 bibs only per photo) | ~80 minutes (batch) + ~90 minutes review of flagged detections |
| Back bib detection accuracy (average distance 10-20m) | 90% for professional riders; 40-60% for amateurs | 55-70% | 94-97% |
| Multi-rider per-photo accuracy (all visible bibs detected) | 20-40 riders per photo is impossible to identify all manually | 1-2 riders max | 40+ riders detected (89-94% confidence on each) |
| Distance handling (bibs at 5m vs 30m away in same frame) | Close bibs readable; distant bibs guessed or skipped | Fails on distance variation; accuracy drops below 40% | Maintains 90%+ accuracy across full distance range |
| Cost per 3000-photo stage | €600-1200 (labor for cross-referencing) | €30-50 (compute) | €90-150 (tokens) |
Shoot from the motorcycle chase position (perpendicular to road, alongside peloton) for maximum bib visibility and consistent distance
Back bibs are most readable when at 90° angle to camera. Shooting from the side-follow position (motorcycle pace alongside) gives consistent distances and angle. Rear-follow position foreshortens bibs; front-ahead position shows no bibs at all.
Capture multiple frames as the peloton moves to get different layers and positions of the same riders
A 10-second motorcycle pass at 60fps equivalent creates 20+ different perspectives of the same riders. Different frames will have different bibs visible (some riders on left in frame A, different riders visible in frame B). AI reads across all frames — you don't need perfection in each individual image.
For tight breakaway groups, position perpendicular to the road at mid-distance to maximize bib legibility across the group
5-8 rider breakaways are easiest to tag from ground-level positions 5-15m from the road, perpendicular to travel direction. This angle gives consistent bib visibility across all riders without extreme foreshortening.
Process stage photos by peloton composition (breakaway, lead group, main field) — review confidence separately for each
Breakaways (5-8 riders) get 95%+ confidence bibs and need <1% manual review. Lead groups (15-25 riders) get 90-94% and need 5-8% review. Main field (40+ riders) gets 85-90% and needs 10-15% review. Allocate review effort per group type.
Cross-reference frame numbers and helmet stickers as secondary ID verification for ambiguous back bibs
Modern cycling bibs are just one ID signal. Pro cyclists also have frame number stickers and helmet branding. If a back bib reads as both #45 and #48 with 87% confidence each, check if the frame number confirms one reading. AI learns to weight multiple ID sources together.
Free trial: upload 50-100 of your peloton photos from any cyclocross, gran fondo, or pro stage. See all riders detected and tagged instantly.
Try cycling peloton tagging →Does back bib detection work when riders are facing at angles (not perpendicular to camera)?
Yes, with reduced accuracy. At 45° angles, bibs are compressed visually and accuracy drops to 85-90%. At extreme angles (near 180°, rear view), bibs distort and confidence flags them for review. The best accuracy is perpendicular (90° angle), but the AI maintains acceptable accuracy across ±45° range.
If a saddle bag completely covers the bib, can the AI still detect it?
Partial coverage is fine; complete coverage is impossible — no AI can read what isn't visible. If 70%+ of the bib is obscured, the AI flags it as low confidence for review. For amateurs with large frame packs that cover bibs, expect 15-20% of riders in photos to be undetectable this way.
Does the system distinguish between frame numbers and back bibs, or does it read both as identification?
RaceTagger focuses on back bibs as the primary ID source. Frame numbers are secondary but the system can read them. For maximum accuracy, import a starting list with both bib→name and frame_number→name mappings. The system will use whichever ID is clearest in each photo.
For time trial events where riders are spaced far apart, is multi-detection still valuable?
Time trials are actually easier — single-rider per frame, no peloton density. Multi-detection accuracy is highest for time trials (98%+) because bibs aren't overlapping. However, the real value of multi-detection is peloton/mass-start events. For TTs, even basic single-subject detection is effective.