Table 1 Works covered in “Solving imbalance within LSL” section
Proposed model | Application | Imbalance type | \(\rho\) | Improvement vs. best baseline | Baseline(s) |
|---|---|---|---|---|---|
No proposed model [16] | LSL image classification | Task, traditional | \(\rho \le 20\) | N/A (survey paper) | Many LSL SOTAs |
\(\alpha\)-TIM [38] | TLSL image classification | Query set | UNSP | + 0–3% Acc. | Many LSL SOTAs |
ProtoNet combined [39]\(^{a}\) | TLSL image classification | Query set | UNSP | + 2% Acc. 1-shot − 1.5% Acc. 5-shot | Many LSL SOTAs + [38] |
TF-MC [40]\(^{a}\) | TLSL image classification | Prediction | UNSP | + 2% Acc. 10-shot | Many LSL SOTAs |
N/A [41]\(^{a}\) | LSL image classification | Traditional | UNSP | + 2% Acc. | Other proposed variants |
N/A [42]\(^{a}\) | LSL image classification | Traditional | \(\rho \le 1000\) | + 2–14% Acc. few-shot + 10–12% F1 Score | Base model + other proposals |
PcGAN [43] | OSL road object classification | Traditional | UNSP | + 5% Acc. | Some OSL SOTAs |
Post-Scaling [44] | Class-incremental image classification | Traditional | \(\rho > 100\) | \(\pm 1\%\) Acc. | Some SOTAs + CI measures |
N/A [45] | LSL semi-supervised image classification | Other | \(\rho = 1\) | + 10–20% Acc. | One non-SOTA model |
MAMC-Net [46] | ZSL domain generalization | Traditional | UNSP | + 0.8% per-class Acc. | Some ZSDG/ ZSL SOTAs |
SCILM [47] | GZSL image classification | Traditional | \(\rho \approx 15\) | + 10% Harm. Mean Acc. | Many ZSL SOTAs |
\(\mathcal {L}_BT\) + GP [48] | GZSL image classification | Traditional | \(\rho \approx 15\) | + 1.5% Harm. Mean Acc. + 8.5% w.r.t. [47] | Many ZSL SOTAs |
DUET [6] | GZSL image classification | Traditional | \(\rho \approx 15\) | + 1.5% Harm. Mean Acc. + 4% w.r.t. [48] | Many ZSL SOTAs |
No Proposed Model [49] | LSL object detection | Traditional | \(\rho > 200\) | N/A (survey paper) | Some LSL SOTAs |
AGCM [50]\(^{a}\) | LSL object detection | Traditional | \(\rho > 200\) | + 0–5 mAP on novel classes | Many LSL SOTAs |
BFS [51] | LSL object detection | Foreground–background | N/A | + 15 mAP + 5 mAP 10-shot w.r.t. [52] | Some LSL SOTAs |
CIR-FSD [52] | LSL object detection | Foreground–background | N/A | + 4-17 mAP + 7 mAP 3-shot w.r.t. [51] | Some LSL SOTAs |
SSL-ALPNet [53] | LSL image segmentation | Foreground–background | N/A | + 25–50% Dice | Two LSL SOTAs |
AMD-Reg [54] | ZSL sketch-based image retrieval | Traditional | \(\rho = 10; 100\) | +3-5 mAP | Many SBIR and ZSL SOTAs |
GwFReID [55] | Re-identification | Traditional | UNSP | + 6 mAP − 9 forgetting ratio | Many Non-LSL SOTAs |
SiameseCCR [56] | LSL character recognition | Contrastive | N/A | + 13% Top-1 Acc. | Some Non-LSL SOTAs |
N/A [57] | Automated CAPTCHA completion | Support set | UNSP | + 0–2% Acc. 10-shot | Base model w/o improvements |
PRNet [58] | OSL 3D image segmentation | Foreground–background | N/A | + 35% Dice | Three OSL SOTAs |
MRE-Net [59] | LSL 3D image segmentation | Foreground–Background | N/A | + 0-6% Dice | One LSL SOTA |
UniFewMeta [60] | LSL Nat. language processing | Traditional | UNSP | + 0–25% Acc. | Three LSL SOTAs |
MELO [61]\(^{a}\) | Cold-start product recommendation | Task | \(5 \le \rho \le 25\) | − 0–0.07 RMSE − 0.0.06 MAE | Base model w/o improvement |
N/A [62] | LSL industrial fault classification | Support set | \(\rho = 10\) | + 15–25% Acc. | Other proposed variants |
RRPN [63] | LSL industrial fault classification | Traditional | UNSP | + 1–1.5% Acc. | Five LSL SOTAs |