The supplementary material is organized as follows:1 Implementation Details.(Sec. Except for the Object Mask Prompt and Compositional Prompt Tuning designs,7 we follow the default architecture in the original ZSseg paper. The number of part queries is set to 50.8 All the two-stage baselines are trained with AdamW optimizer with the initial learning rate of 1e-49 and weight decay of 1e-4. A poly learning rate policy with a power of 0.9is adopted.

In this document, we provide more implementation details of CATs and more results on SPair71k [16], PF-PASCAL [4], and PF-WILLOW [3]. Given resized input images Is,It R256 256 3, we conducted experiments using different feature backbone networks, including DeiT-B [22], DINO [2] and ResNet-101 [5]. For the ResNet-101multi in the paper, we use the best layer subset [15] of (0,8,20,21,26,28,29,30) for SPair-71k, and (2,17,21,22,25,26,28) for PF-PASCAL and PF-WILLOW. We resized the spatial resolution of extracted feature maps to 16 16. The extracted features undergo l-2 normalization and the correlation maps are constructed using dot products.

As discussed in prior literature [1, 4], one operation of floating-point addition and multiplication have energy costs of 0.9 pJ and 3.7 pJ, respectively. Meanwhile, one operation of 8-bit integer addition and multiplication have 0.03 pJ and 0.2 pJ energy costs, demonstrating much lower cost than floating-point operation. Therefore, it is important to explore whether adder detectors performs well for INT8 quantization. We tried to adopt INT8 post quantization for our Adder FCOS (B+N) model, which suffers 0.8 mAP drop compared with full precision model, as shown in Table A. The energy reduction further increases from 29% to 35%. Note that post training quantization is not optimal for INT8 models, and quantization-aware training may greatly further improve the accuracy.

A.1 Network Architectures We adopt Daformer [17] with Swin-B or MiT-B5 backbone as the base semantic segmentation architecture. For the segmentation head, we utilize the same head as Daformer [17]. The stem module contains one fully-convolutional layers with kernel 3 3 and stride of 2, two fully-convolutional layers with kernel 3 3 and stride of 1, two fully-convolutional layers with kernel 3 3 and stride of 2, and another three fully-convolutional layers with kernel 1 1 and stride of 1 to adjust channels of different feature maps. Level embedding module is defined as metrics with shape 3 dims. The prompt Interactor module contains three fully-convolutional layers with kernel 3 3 and stride of 2 to adjust feature dimensions.