This is the official implementation of the following publication:

AirRoom: Objects Matter in Room Reidentification
Runmao Yao, Yi Du, Zhuoqun Chen, Haoze Zheng, Chen Wang
CVPR 2025
arXiv | Project Page

The code has been tested on:

• CUDA 11.5
• GeForce RTX 3090 Ti (24GB).

1. Clone the repository

   git clone https://github.com/21yrm/AirRoom.git
   cd AirRoom
   

2. Build the Docker container

   docker login
   ./build_docker.sh
   

3. Download the dataset

   Download the room re-identification dataset from the following link: Google Drive – Room Re-identification Dataset

   After downloading, unzip the datasets.zip file and place the extracted folder under the project root directory.

   The expected dataset directory structure should look like:

   datasets/
   ├── <dataset_1>/
   │   ├── <scene_1>/
   │   │   ├── <room_1>/
   │   │   │   ├── rgb/
   │   │   │   └── depth/
   │   │   └── ...
   │   ├── ...
   │   └── room_label.txt
   └── ...
   

Follow the steps below to run the project:

1. Launch the Docker Container

   ./run_docker.sh
   

2. Preprocess the Dataset (Construct Reference Database)

   • Open config/preprocess.yaml

   • Set the field dataset_name to the desired dataset (e.g., MPReID)

   • Then run:

     python preprocess.py
     

3. Fix Compatibility Issues for Python 3.8 and PyTorch 1.x

   Some operations are not supported in Python 3.8 and PyTorch 1.x. Please modify the following files accordingly:

   cache/torchhub/hub/facebookresearch_dinov2_main/dinov2/layers/attention.py

   • Line 58:

     self, init_attn_std: float | None = None, init_proj_std: float | None = None, factor: float = 1.0
     

   • Change to:

     self, init_attn_std = None, init_proj_std = None, factor: float = 1.0
     

   • Before Line 69:

     Insert the following function before the forward method:

     def scaled_dot_product_attention(self, query, key, value, attn_mask=None, dropout_p=0.0,
         is_causal=False, scale=None, enable_gqa=False) -> torch.Tensor:
         import math
         L, S = query.size(-2), key.size(-2)
         scale_factor = 1 / math.sqrt(query.size(-1)) if scale is None else scale
         attn_bias = torch.zeros(L, S, dtype=query.dtype, device=query.device)
         if is_causal:
             assert attn_mask is None
             temp_mask = torch.ones(L, S, dtype=torch.bool).tril(diagonal=0)
             attn_bias.masked_fill_(temp_mask.logical_not(), float("-inf"))
             attn_bias.to(query.dtype)
     
         if attn_mask is not None:
             if attn_mask.dtype == torch.bool:
                 attn_bias.masked_fill_(attn_mask.logical_not(), float("-inf"))
             else:
                 attn_bias = attn_mask + attn_bias
     
         if enable_gqa:
             key = key.repeat_interleave(query.size(-3)//key.size(-3), -3)
             value = value.repeat_interleave(query.size(-3)//value.size(-3), -3)
     
         attn_weight = query @ key.transpose(-2, -1) * scale_factor
         attn_weight += attn_bias
         attn_weight = torch.softmax(attn_weight, dim=-1)
         attn_weight = torch.dropout(attn_weight, dropout_p, train=True)
         return attn_weight @ value
     

   • Lines 102–104:

     x = nn.functional.scaled_dot_product_attention(
         q, k, v, attn_mask=None, dropout_p=self.attn_drop if self.training else 0, is_causal=is_causal
     )
     

   • Change to:

     x = self.scaled_dot_product_attention(
         q, k, v, attn_mask=None, dropout_p=self.attn_drop if self.training else 0, is_causal=is_causal
     )
     

   cache/torchhub/hub/facebookresearch_dinov2_main/dinov2/layers/block.py

   • Lines 150–152:

     init_attn_std: float | None = None,
     init_proj_std: float | None = None,
     init_fc_std: float | None = None,
     

   • Change to:

     init_attn_std = None,
     init_proj_std = None,
     init_fc_std = None,
     

   Then re-run:

   python preprocess.py
   

4. Run Inference

   • Modify config/inference.yaml to specify the target dataset and desired configuration.

   • Then execute:

     python inference.py
     

If you find our work interesting, please consider citing us!

@InProceedings{Yao_2025_CVPR,
    author    = {Yao, Runmao and Du, Yi and Chen, Zhuoqun and Zheng, Haoze and Wang, Chen},
    title     = {AirRoom: Objects Matter in Room Reidentification},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {1385-1394}
}

We gratefully acknowledge the following open-source projects that significantly contributed to our work:

• Semantic-SAM, for its outstanding instance segmentation performance.

• LightGlue, for its excellent local feature matching capabilities.