Add Salad retrieval

hloc/extract_features.py

@@ -146,6 +146,11 @@
         "model": {"name": "eigenplaces"},
         "preprocessing": {"resize_max": 1024},
     },
+    "salad": {
+        "output": "global-feats-salad",
+        "model": {"name": "salad"},
+        "preprocessing": {"resize_max": 640},
+    },
 }
 
 

hloc/extractors/salad.py

@@ -0,0 +1,45 @@
+"""Global image feature extractor Salad:
+Optimal Transport Aggregation for Visual Place Recognition
+Sergio Izquierdo, Javier Civera; CVPR 2024.
+https://github.com/serizba/salad
+"""
+import math
+
+import torch
+import torchvision.transforms as tvf
+
+from ..utils.base_model import BaseModel
+
+
+class Salad(BaseModel):
+    default_conf = {
+        "backbone": "dinov2_vitb14",
+        "pretrained": True,
+        "patch_size": 14,
+    }
+    required_inputs = ["image"]
+
+    def _init(self, conf):
+        self.net = torch.hub.load(
+            "sarlinpe/salad",
+            "dinov2_salad",
+            backbone=conf["backbone"],
+            pretrained=conf["pretrained"],
+        ).eval()
+
+        mean = [0.485, 0.456, 0.406]
+        std = [0.229, 0.224, 0.225]
+        self.norm_rgb = tvf.Normalize(mean=mean, std=std)
+
+    def _forward(self, data):
+        image = self.norm_rgb(data["image"])
+        _, _, h, w = image.shape
+        patch_size = self.conf["patch_size"]
+        if h % patch_size or w % patch_size:
+            h_inp = math.ceil(h / patch_size) * patch_size
+            w_inp = math.ceil(w / patch_size) * patch_size
+            image = torch.nn.functional.pad(image, [0, w_inp - w, 0, h_inp - h])
+        desc = self.net(image)
+        return {
+            "global_descriptor": desc,
+        }