Indoor Object Detection — YOLO11 Baseline

A Google Colab YOLO11s baseline on the Indoor Object Detection Dataset, using a strict sequence-aware split to avoid adjacent-frame leakage while reporting overall and per-class validation mAP for all seven classes.

Open primary Colab notebook View public repository Checkpoint release

Result summary

Item	Value
Model	YOLO11s pretrained detector
Validation mAP@0.5	0.9142
Validation mAP@0.5:0.95	0.6797
Validation split	sequence 2, 217 images, 346 boxes
Training runtime	Google Colab Tesla T4, 75 epochs, batch 16, image size 640

Split and leakage control

The dataset images come from six source video sequences. Random image-level splitting can place adjacent or near-duplicate frames across train/validation/test. I used whole source sequences instead.

Split	Source sequences	Images	Percentage
train	1, 3, 4, 5	1,809	81.744%
validation	2	217	9.806%
test	6	187	8.450%

The split is approximate rather than exact 80/10/10 because sequence-disjoint evaluation is more reliable than exact percentages. All seven classes appear in every subset.

Split support summary — Per-class support in train, validation, and test.

Validation metrics

Metric	Value
precision	0.9255
recall	0.8228
mAP@0.5	0.9142
mAP@0.75	0.7892
mAP@0.5:0.95	0.6797

Class	Boxes / images	Precision	Recall	mAP@0.5	mAP@0.5:0.95
chair	127 / 70	0.8820	0.4016	0.6808	0.3613
clock	25 / 25	1.0000	0.9963	0.9950	0.8080
exit	59 / 59	0.9293	0.9661	0.9891	0.8136
fireextinguisher	53 / 51	1.0000	0.9807	0.9950	0.7633
printer	23 / 23	0.9024	1.0000	0.9950	0.8211
screen	8 / 8	0.8120	0.6250	0.7811	0.5122
trashbin	51 / 50	0.9527	0.7901	0.9637	0.6787

Training curves from the accepted baseline run.

Per-class validation metrics — Per-class validation mAP@0.5:0.95 with support.

Good and bad validation examples

Good validation examples — Validation predictions with class and confidence labels.

Validation ground-truth and prediction overlay — Validation ground truth and predictions: green is ground truth, orange is prediction.

Bad validation examples — Failure-oriented validation examples with dense chair clusters, misses, and false positives.

Limitations

The result is intentionally a single baseline rather than a tuned benchmark. Chair is the weakest class because chairs often appear in dense overlapping groups with occlusion, truncation, and varied viewpoints. Screen performance is also uncertain because the validation split has only eight screen boxes.

Reproducibility

The primary Colab notebook is the reproducibility path. It opens in a fast review mode by default, can download the official dataset for live dataset/split checks, and can optionally download the published checkpoint release asset to rerun validation. The checkpoint is a GitHub Release asset, not a git-tracked file.

Release: v1.0-submission

AI-assisted workflow

AI agents accelerated dataset auditing, split validation, Colab execution, artifact handling, and notebook QA. Final ML/evaluation decisions were manually reviewed. See AI_ASSISTED_WORKFLOW.md.