Multi-Point Positional Insertion Tuning for Small Object Detection

TL;DR

The paper tackles small object detection by proposing MPI tuning, a parameter-efficient finetuning method that inserts multi-point positional embeddings into a frozen pretrained detector. The core is a lightweight MHP encoder consisting of sinusoidal embeddings, tiny MLPs, and a multi-head mixer that produces per-feature positional vectors $m{p}_i$, added to latent features via $m{h}'_i(m{x}) = m{h}_i(m{x}) + m{p}_i$. Applied to Grounding DINO and evaluated on SODA-D, MPI achieves competitive mAP with only 0.50M trainable parameters, significantly reducing finetuning cost while preserving performance relative to CoOp and VPT; ablations confirm the importance of injecting positional information into the feature enhancer. The work advances practical, resource-efficient small-object detection in vision-language pretrained frameworks and points to future gains from refined insertion strategies and backbone adaptations.

Abstract

Small object detection aims to localize and classify small objects within images. With recent advances in large-scale vision-language pretraining, finetuning pretrained object detection models has emerged as a promising approach. However, finetuning large models is computationally and memory expensive. To address this issue, this paper introduces multi-point positional insertion (MPI) tuning, a parameter-efficient finetuning (PEFT) method for small object detection. Specifically, MPI incorporates multiple positional embeddings into a frozen pretrained model, enabling the efficient detection of small objects by providing precise positional information to latent features. Through experiments, we demonstrated the effectiveness of the proposed method on the SODA-D dataset. MPI performed comparably to conventional PEFT methods, including CoOp and VPT, while significantly reducing the number of parameters that need to be tuned.

Figures (4)

• Figure 1: Multi-point positional insertion (MPI) tuning for small object detection. MPI tuning inserts positional embeddings at multiple points in a frozen pretrained model through a learnable multi-head positional encoder (MHPE). This figure illustrates a frozen object detection model with $N$ sequential modules for simplicity.
• Figure 2: (a) Multi-head positional encoder consisting of sinusoidal positional embeddings, tiny MLPs, and a multi-head mixer. (b) Architecture of each tiny MLP.
• Figure 3: Application to GDINO. The points to insert embeddings $\bm{p}_{i}$ are highlighted in green. (a) Architecture of GDINO. (b) Architecture of feature enhancer block. (c) Architecture of decoder block.
• Figure 4: Qualitative examples