Discord

Checklist AICC 2026 Round 5 · Task 3

Visual Relations

Predict ranked (subject box, predicate, object box) relationship triplets in images using only a DETR detector.

  • Vision
  • Visual relationship detection
Read the official statement

The task

Smart glasses for visually impaired people must not only detect objects but also describe how they relate, for example whether a person is holding a cup or merely standing near it.

The training set contains 4000 images with about 15k annotated relationships (3.75 per image on average), each giving subject and object boxes (pixel xyxy), their class names and a predicate from 35 classes (spatial relations, actions and interactions such as riding, holding, on, near, wearing). The test set contains 1000 images. Every annotated training subject and object is detectable by the provided model at IoU >= 0.5, and a pair may carry several predicates.

For each test image the contestant outputs a confidence-ranked list of triplets: subject box, predicate and object box.

Abridged by SOTA from the official materials. The official statement has the exact rules, and it wins wherever this summary differs.

At a glance

You get
images/ (4000 training and 1000 test JPEG images), train_annotations.csv, train_images.csv, test_images.csv and predicates.json (ordered list of 35 predicates).
You submit
submission.csv with columns image_id and PredictionString; each prediction is 10 space-separated tokens 'predicate confidence sub_x1 sub_y1 sub_x2 sub_y2 obj_x1 obj_y1 obj_x2 obj_y2'; at most 100 predictions per image; every test image must appear (empty string if none).
Scoring
Mean Average Precision over the 35 predicate classes with equal weight. Per class, predictions are pooled and sorted by confidence; a prediction is a true positive if the predicate matches and both subject and object boxes have IoU >= 0.5 with an unmatched ground-truth triplet (greedy matching by best IoU sum); AP uses 11-point interpolation. Malformed submissions score 0. Baseline 0.0029, reference solution 0.1769.
Rules
  • facebook/detr-resnet-50 is the only pretrained model allowed.
  • No external data beyond what is provided.
  • Individual participation (maximum team size 1); at most 15 submissions per day.
  • AICC contest rules (stated on each Kaggle rules page, not enforceable): no use of LLMs for writing code or getting task ideas; no internet use other than official library documentation and the contest platform; no communication with anyone during the contest; clarifications only via the #clarification-requests channel on the AICC Discord server.
Format
AICC Round 5, online on Kaggle, 27 Mar 2026 18:00 UTC – 29 Mar 2026 18:00 UTC.

Notebook

Competition

Solutions

Details

Year
2026, Online (Kaggle)
Round
Round 5 · Task 3
Language
English
License
Varies by task: Watermark Removal and Visual Relations — CC BY-NC-SA 4.0; Lost Interpreter — Apache 2.0 (Kaggle competition licences). Solutions repository: MIT., as stated by the source