Ten active-learning methods for object detection sit within one mAP50 point of random selection on MS-COCO

What changed, and when

Licheng Zhang (University of Melbourne) and Zheng Gong (Jimei University) submitted Distance to Class Prototypes: Active Learning for Object Detection on 28 July 2026; arXiv announced it as 2609.20248v1 in its September 2026 listings. The proposal: add a supervised contrastive branch to an SSD detector so that each detected box lands in a 1,024-dimensional space where distance means class membership, then score unlabelled images by how far their detections sit from the anchors of their predicted class, weighted by confidence.

The pitch is cost. The branch adds 2.89M parameters (8.3% on VOC, 7.2% on COCO), needs one forward pass (0.0037 s against 0.0032 s for the bare detector, plus 0.008 s to score against 1,000 anchors), and no ensemble. The paper cites prior work noting that one detector can cost up to 282 GPU hours to train, which is what makes three-model ensembles and 25 to 50 MC-dropout passes hard to afford inside an annotation campaign.

The protocol is the field's standard one: SSD with VGG-16, 500 epochs, retrained from ImageNet each cycle, 1,000 images added per cycle; VOC07 from 2,000 to 4,000 images, VOC07+12 from 1,000 over nine cycles, MS-COCO from 5,000 to 7,000. Random and the authors' method were run here with three seeds. Every other row is quoted from earlier papers.

On MS-COCO the field fits inside one mAP50 point

Three numbers from Table 4 frame everything else. The best method beats random by 0.84 points at 7,000 images. Two published methods, LLAL and Feature-mixture, sit below random. And the 5,000-image column, where no selection has happened and every method trains on a random initial set, already spans 0.61 points (27.10 to 27.71). That first-round spread is implementation: different codebases, schedules and seeds reaching different scores on equivalent data. It is about two thirds of the whole method-to-method spread after two rounds of selection.

VOC07 is kinder to active learning. At 4,000 images the best method (Feature-mixture, 69.98) is 1.59 points over random (68.39) and every method is above random. The initial-round spread there is 0.54 points. The authors' method reaches 69.78, 0.20 behind Feature-mixture; on COCO the ordering reverses and Feature-mixture (29.60) is below random. Feature-mixture is the first author's own earlier method, which the paper discloses in Section 2.4.

The paper is candid about the comparison it can actually defend. Its claim is against Entropy, the class posterior of the same detector under the same protocol: +0.92 and +1.08 points on VOC07 at 3k and 4k, +0.07 and +0.62 on COCO at 6k and 7k, against run-to-run deviations of 0.02 to 0.18. Three of four gaps are an order of magnitude above the noise. The authors also note the gap "carries the value of the signal together with whatever implementation difference separates the two", because Entropy's numbers come from another codebase.

How the score is computed, and what it costs

The ablation (Table 6) carries the mechanistic claim. Apply the same distance rule to embeddings from a detector trained without the contrastive loss and COCO comes out at 28.77 and 29.57, next to random's 28.67 and 29.67: the shaped space is doing the work, with the largest single gap 0.94 points at 7k. Dropping the confidence factor from the score costs 1.35 to 2.38 points in every one of six VOC07+12 rounds (Figure 5), so the geometry adds to the posterior rather than replacing it. Swapping in a scoring rule from image-classification active learning costs only 0.03 to 0.29. On COCO, the paper's 42.77M parameters are a little over a third of the 116.51M most rows carry and an eighth of the ensemble's 349.53M.

A labelling budget is counted in images, not mAP, so we read the gains across the random curve. On VOC07 the method's 67.77 at 3,000 images is where random selection would be at about 3,698 images: roughly 698 images, or 23% of the labelled set, saved. On COCO its 29.00 at 6,000 matches random at about 6,330, 330 images or 5.5%. Linear interpolation understates concave curves slightly, and the 4k and 7k points sit past random's last measured round, so we give no figure there.

What it means for a team deciding which images to label

• Run random as your own arm, in your own code, every time. Here the published leaderboard mixes three-seed reruns with numbers quoted from other papers, and the pre-selection spread (0.61 on COCO) is most of the post-selection spread. A method that cannot beat your random arm by more than your own first-round spread has not been shown to work on your data.
• Price the gain in labels. On 80-class COCO with 1,000-image cycles, the best case here is a few hundred images per cycle, about 6% of the set. On 20-class VOC it is about 23%. Multiply by your per-image box-annotation cost to decide whether any selection machinery is worth building.
• If you do use it, prefer single-pass signals. At 7k on COCO the prototype-distance score matched or beat ensembles and MC-dropout (at 6k it trailed both by 0.03 and 0.20) with a third to an eighth of the parameters and one pass. The expensive signals are not buying accuracy at this scale.
• Expect fine-grained, many-class sets to be harder. Every published method beat random on VOC07; two did not on COCO. Our reading, not tested in the paper: with 80 classes a 5,000-image start covers each class thinly, so any score built on the model's own beliefs is scoring a weak model.

Judgement: this is a careful paper with a modest, honestly stated claim, and it is a good reference for how small the active-learning margin in detection is once you line the numbers up. The same pattern appeared in BioDCASE 2026's bioacoustics challenge, where two of fourteen ranked methods also finished below random.

Check it yourself

No code link in v1. Everything above is arithmetic on Tables 3 and 4:

python3 - <<'EOF'
coco = {"Random": (27.17, 28.67, 29.67), "Entropy": (27.70, 28.93, 29.89), "Core-set": (27.70, 28.99, 29.93),
        "LLAL": (27.71, 28.71, 29.53), "Prob": (27.33, 29.06, 30.02), "GMM": (27.70, 29.28, 30.51),
        "Feature-mixture": (27.10, 28.60, 29.60), "MC-dropout": (27.70, 29.20, 30.30),
        "Ensemble": (27.70, 29.03, 30.02), "Ours": (27.33, 29.00, 30.51)}          # Table 4, mAP50 at 5k/6k/7k
r = coco["Random"][2]
print("spread at 7k:", round(max(v[2] for v in coco.values()) - min(v[2] for v in coco.values()), 2))
print("spread at 5k:", round(max(v[0] for v in coco.values()) - min(v[0] for v in coco.values()), 2))
print("below random at 7k:", [k for k, v in coco.items() if v[2] < r])
def interp(score, pts):
    for (n0, s0), (n1, s1) in zip(pts, pts[1:]):
        if s0 <= score <= s1: return round(n0 + (score - s0) / (s1 - s0) * (n1 - n0))
print("VOC07 67.77 =", interp(67.77, [(2000, 62.37), (3000, 66.34), (4000, 68.39)]), "random images")
print("COCO 29.00 =", interp(29.00, [(5000, 27.17), (6000, 28.67), (7000, 29.67)]), "random images")
EOF
# expected: 0.98 / 0.61 / ['LLAL', 'Feature-mixture'] / 3698 / 6330

• Protocol to reproduce the arms: SSD300, VGG-16 from ImageNet, 500 epochs, batch 24, 50 warm-up epochs from 1e-5 to 1e-3, retrain from scratch each cycle, 1,000 images per cycle, VOC07 11-point mAP50 and COCO-style mAP50 (Section 4.2, Table 2).
• The VOC07+12 comparison (nine cycles, Figure 4) is reported only as curves; the authors state it supports ordinal comparisons only.

What would prove this wrong

Prediction, dated: by 30 June 2027, no detection active-learning paper that reruns random and at least three published baselines in one codebase on MS-COCO (5,000 initial images, 1,000 per cycle, two or more cycles) will report a gain over random above 1.5 mAP50 points at the second selection round. A single-codebase result above 1.5 would mean the tight spread here is an artefact of mixing quoted numbers and the selection signal matters more than this table suggests.

Sources

1. Licheng Zhang and Zheng Gong, Distance to Class Prototypes: Active Learning for Object Detection, arXiv:2609.20248v1, submitted 28 July 2026, announced September 2026. Tables 1 to 6, Sections 3 and 4, Figure 4.
2. BLOMEGA, BioDCASE 2026 Task 4: active learning for bioacoustics, lab note on arXiv:2609.15255 (results published 14 September 2026).

Related BLOMEGA research: Annotation budget between SFT and RL · Interactive annotation of radiology volumes · Data annotation: latest research

FAQ

Does active learning beat random selection for object detection?

Barely, on the standard benchmark. In arXiv:2609.20248 (September 2026), ten methods on MS-COCO at 7,000 labelled images span 29.53 to 30.51 mAP50 against random's 29.67; two (LLAL and Feature-mixture) are below random. On PASCAL VOC07 at 4,000 images every method beats random, by up to 1.59 points.

How many labels does active learning save in object detection?

Reading the paper's numbers across the random-selection curve: about 698 images at 3,000 on VOC07 (23%) and about 330 at 6,000 on MS-COCO (5.5%) for the prototype-distance method.

What is prototype-distance active learning?

A supervised contrastive branch (2.89M parameters, 8.3% extra on SSD) embeds each detected box so distance encodes class. An unlabelled image is scored by how far its detections sit from 1,000 anchors of their predicted class, weighted by confidence, in a single forward pass.

Why do active learning leaderboards in detection mislead?

Most rows are quoted from other codebases. In this paper's MS-COCO table the methods already differ by 0.61 mAP50 points before any selection, on random initial sets, which is about two thirds of the 0.98-point spread after two rounds.