SudhishNeel/clash-royale-detection

Clash Royale arena detection

Labelled Clash Royale arena data for detecting units, buildings, spells and towers, built for a Hog 2.6 playing bot. Fan-made research data, not affiliated with or endorsed by Supercell. Last updated 2026-10-15.

Part	What	Size
crops config (default)	the author's graded unit crops, one per graded track	27,500 crops (18,995 by hand + 8,505 auto-filled)
crops_predicted config	YouTube track crops with the crop classifier's call: confidence ≥ 0.7 or approved by the author	grows as the author approves more
track_grades/	the author's per-track grades with every box	99 videos
track_labels/<video_id>/	per-video label files: author's answers, auto-fills, rejections, trusted training labels	99 videos
synthetic_arena config	real-arena synthetic frames: real YouTube arena backdrops + real unit cut-outs, with boxes	105,000 frames, 158 classes + spells
archive/synthetic/	archived: 6,000 synthetic frames with boxes (pasted cut-outs, too rough to use)	6,000 frames
video_labels/	YOLO labels for broadcast / YouTube frames (labels only)	11,107 files, 6 videos
bundles/	everything the labelling pipeline made per video	75 videos, 4.1 GB

Frames from the author's own phone games and Training Camp screenshots are not included.

Format

• Frame: the 568x896 arena crop (18x32 tile grid; queen towers at (116,200) (461,200) (116,689) (461,689), kings at (284,108) (284,790)), JPEG. See SPEC.md.
• Archived synthetic parquet: objects.bbox is [x0, y0, x1, y1] in pixels; objects.category uses the 208 class names below.
• Class list: classes.txt / data.yaml. The class order is permanent: new classes are only appended, so old label files stay valid.

Classes (208)

Every card, evolution and hero as of 8 Oct 2026. A class is a unit as it appears on the field, not a card: Skeleton Army = skeleton, Goblin Gang = goblin + spear-goblin, Minion Horde = minion, Three Musketeers = musketeer.

• Evolution: diamond level badge, class <card>-evolution.
• Hero: crown badge, its own class <card>-hero (since 2026-10-08), e.g. musketeer-hero, ice-wizard-hero, giant-hero. Champions (golden-knight, skeleton-king, archer-queen, mighty-miner, monk, little-prince, goblinstein, boss-bandit) have their own class.
• Spells are classes: poison, earthquake, graveyard, rage, freeze, arrows, the-log, barbarian-barrel and so on label the spell's area or projectile. Fireball explosions and tornado swirls are not labelled.
• No examples yet:
  • terry: a removed one-time card.
  • zap and zap-evolution: too quick to see.
  • rocket, vines, heal-spirit: parked, because in gameplay footage they are a blur, a ground effect or tiny, and too hard to box reliably.
• Pending splits: these are labelled as the base card for now. bowler vs bowler-hero; barbarian-barrel vs barbarian-barrel-hero (hero = golden spikes, normal = black spikes); minion vs minion-evolution.

Full rules are in LABELLING.md.

How the video labels are made

YouTube gameplay (pro broadcasts, top-ladder POV, card guides) is normalised to the arena frame. Then:

1. Tracks. A detector finds the units and they are tracked across frames, so one track is one unit.

2. Hand grading. The author names tracks in a grading tool: 18,995 tracks by hand so far.

3. Auto-fill. Each answer is copied to near-identical tracks of the same video (DINOv2 crop similarity ≥ 0.8, 96-100% right when checked): 8,505 more.

4. Crop classifier. A crop classifier is trained on those answers (see the model repo SudhishNeel/clash-royale-detector, crop_classifier/). Its confident calls were checked yes/no by the author.

5. Trusted labels only. A track is used for training only if one of these is true; everything else is painted out of the frame:

   • the author named it, or it was auto-filled from the author's answer;
   • an agent label agrees with the classifier (audit: 38/39 right);
   • the classifier alone is at confidence ≥ 0.85, except for goblin / spear-goblin (tightened after the audit: at ≥ 0.7 these calls were only 28/40 right, goblins 1/8);
   • for spells, the detector's spell call is in the classifier's top 3 (audit: 18/20 right).

   Audit = the author's yes/no check of a random sample of machine-trusted tracks (2026-10-09).

   Tracks the author rejected are never trusted. In hero videos, base-card labels are not trusted unless the author gave them, so bowler is never confused with bowler-hero.

Status of the trusted labels (2026-10-09; the trusted files are refreshed at the next label upload):

• 46,919 tracks in 65 graded videos; 26,496 unit tracks are trusted.
• Where the trusted labels come from: 14,038 by the author, 7,340 auto-filled, 6,237 agreement, 1,420 from the classifier alone, 416 spells.
• 17,739 tracks are left out.
• New footage is being added for the thinnest classes: rocket, goblin-barrel, void, vines, heal-spirit, boss-bandit, goblin-machine, wizard-hero and bowler-hero.

Agent-only labels (LLM agents naming crops) agreed with the author only about 60% of the time. That is why none are used unless the classifier agrees.

Hand grading (track_grades/)

Per-track class grades from the dataset author for YouTube gameplay videos. No frames are stored. track_grades/<video_id>/grades.csv has one row per box of every graded track, with these columns: track_id, label, source, frame, is_best, x0, y0, x1, y1. Coordinates are pixels on the 568x896 frame.

• source is user (graded by hand) or auto (the author's answer copied to a near-identical track).
• label can be drop, meaning not a unit (effects, UI, text).
• is_best marks the frame the grade was made on. Other frames of a track can drift onto a neighbouring unit.

To rebuild the frames, run python scripts/rebuild_video_frames.py <video_id> video.mp4 out_dir track_grades on a video you have downloaded yourself.

Graded crops (crops config, shown by default)

One row per graded track: the unit cut out of its box on the frame it was graded on. The columns are image, label, source, video_id, track_id, frame.

• label is one of the 208 classes or drop.
• video_id, track_id and frame link back to track_grades/.

Real-arena synthetic frames (synthetic_arena config)

Extra training frames for the rarer cards, built only from real gameplay pixels (2026-10-15):

• Backdrop: a real arena from a YouTube match, cleaned by taking the per-pixel median over that match's frames with every unit masked out (113 arenas).
• Units: real units cut out of the same footage at their true in-game scale, from tracks the author labelled (or the crop classifier confirmed). Every cut-out was checked by the crop classifier on its own picture, so deploy clouds, clocks and name captions are left out. 702 cut-outs over 129 classes.
• Placement: game-plausible: troops in lanes, on bridges and around towers, buildings in their owner's half, no ground units on the river. 36-40% of frames have overlapping fights (back units partly hidden); a unit is only pasted and labelled if at least 30% of it stays visible.
• Balance: rarer classes appear more often (every class at least 461 times).
• Three batches:
  • train-0000*: 20,000 frames, 129 classes, the strictest cut-out check.
  • train-b2-* (ids synth_b2_*): 6,000 frames aimed at the rarer cards, 158 classes, about 80% of units from cards with few cut-outs. For tracks the author labelled, each cut-out is compared with the author's own graded crops of that card instead of requiring the crop classifier, which was trained before many rare cards were labelled.
  • train-b3-* (ids synth_b3_*): 4,000 frames with 1-3 spells each, from the author's hand-labelled spells. Area spells (rage, poison, freeze, graveyard, earthquake, arrows, void, goblin-curse) are pasted as see-through overlays under the units, so the arena shows through; the-log, barrels, snowball and axe are pasted like units.
• Bulk batches: train-c###* (ids synth_c###*): runs of 5,000 frames, each with its own randomly drawn settings: units per frame (2-20), share of overlapping fights (15-60%), spells in about 60% of runs (0-4 per frame) and rare-card targeting in about 30% of runs. Same icon-free cut-outs and spell overlays as batches 2 and 3.
• No icons: cut-outs showing a deploy clock, an emote, a caption or another overlay icon were found by a visual review and left out of batches 2 and 3 (138 cut-outs). Batch 1 was made before that review and still has some.
• Not in the synthetic frames yet: boss-bandit, dirt, royal-delivery, tornado and a few evolutions. They do have real hand-graded labels (see crops); no clean cut-out of them could be made yet.

Columns: image (568x896), image_id, backdrop_video (the match the arena came from), n_boxes, objects.bbox ([x0, y0, x1, y1] pixels) and objects.category (the 208 classes). The author checked a random sample by eye before upload. Use these alongside the real labels, not instead of them.

Classifier-labelled crops (crops_predicted config)

YouTube track crops (112px tight crops) with the crop classifier's call, where the call is confident or the author approved it. A row is included when either holds:

• the classifier's confidence is at least 0.7;
• the author graded or approved the track: lower-confidence calls are checked one by one in a yes/no tool, and only approved ones are added.

Tracks the author rejected are never included.

Column	Meaning
video_id, track_id	the track
pred, conf, top3	the crop classifier's call (fine-tuned on 24,421 graded tracks)
trusted	the label used for training; empty means left out (see "How the video labels are made")
graded	user or auto when the author graded it; those rows are also in crops

Per-video label files (track_labels/<video_id>/)

The current label files of every graded video: names_user.json (the author's answers), auto_user.json (auto-fills), yn_rejects.json (classifier calls the author rejected), trusted.json + trusted_stats.json (the filtered training labels). Newer than the copies inside bundles/.

Raw bundles (bundles/<video_id>.tar)

Everything the labelling pipeline made for one video:

• tracks.json: every track's boxes per frame.
• crops/ and ctx/: track crops and whole-arena previews.
• decks/ and layout.json.
• Every label file:
  • agent passes: names_*, unsure_*;
  • Opus review: review_*, with *.pre_opus.json holding the labels from before the review;
  • the author's answers: names_user.json, auto_user.json, and yn_rejects.json where present (classifier calls the author rejected);
  • the filtered training labels: trusted.json, with trusted_stats.json.
• DINOv2 tight-crop embeddings: emb_tight.npz.