changes so multi opj works in image 77, need to test on others

main_lost.py

@@ -306,17 +306,17 @@ if __name__ == "__main__":
             elif args.visualize == "seed_expansion":
                 for i, x in enumerate(zip(preds, seeds)):
                     pred, seed = x
-                image = dataset.load_image(im_name)
-
-                # Before expansion
-                pred_seed, _ = detect_box(
-                    A[seed, :],
-                    seed,
-                    [w_featmap, h_featmap],
-                    scales=scales,
-                    initial_im_size=init_image_size[1:],
-                )
-                visualize_seed_expansion(image, pred, seed, pred_seed, scales, [w_featmap, h_featmap], vis_folder, im_name+'_'+str(i))
+                    image = dataset.load_image(im_name)
+
+                    # Before expansion
+                    pred_seed, _ = detect_box(
+                        A[seed, :],
+                        seed,
+                        [w_featmap, h_featmap],
+                        scales=scales,
+                        initial_im_size=init_image_size[1:],
+                    )
+                    visualize_seed_expansion(image, pred, seed, pred_seed, scales, [w_featmap, h_featmap], vis_folder, im_name+'_'+str(i))
 
             elif args.visualize == "pred":
                 image = dataset.load_image(im_name)
@@ -340,9 +340,8 @@ if __name__ == "__main__":
                 continue
 
             if len(gt_bbxs) == 0:
-                continue # TODO: should do something else, should skip iou but count towards FP if pred exists
+                break # TODO: should do something else, should skip iou but count towards FP if pred exists
 
-            # TODO: preds needs to be
             ious = bbox_iou(torch.from_numpy(pred), torch.from_numpy(np.asarray(gt_bbxs)))
 
             # TODO: This calculates the corloc

object_discovery.py

@@ -56,41 +56,58 @@ def lost(feats, dims, scales, init_image_size, k_patches=100, num_init_seeds=1,
     """
     # Compute the similarity
     A = (feats @ feats.transpose(1, 2)).squeeze()
-
-    if num_init_seeds== 0:
-        num_init_seeds = len(A)
-
     # Compute the inverse degree centrality measure per patch
     sorted_patches, scores = patch_scoring(A)
 
-    # Select the initial seed
-    seeds = sorted_patches[0:num_init_seeds]
+    if num_init_seeds <= 0:
+        num_init_seeds = len(sorted_patches)
 
     preds = []
     filtered_seeds= []
 
+    # if only one seed, then do default LOST and return
+    if num_init_seeds == 1:
+        seed = sorted_patches[0]
+        potentials = sorted_patches[:k_patches]
+        similars = potentials[A[seed, potentials] > 0.0]
+        M = torch.sum(A[similars, :], dim=0)
+        pred, _ = detect_box(M, seed, dims, scales=scales, initial_im_size=init_image_size[1:])
+        pred = np.asarray(pred)
+        filtered_seeds.append(seed)
+        preds.append(pred)
+        return np.asarray(preds), A, scores, filtered_seeds
+    
+    #print('feats.shape',feats.shape)
+
     # Seed expansion
-    for i, seed in enumerate(seeds):
+    for i in range(num_init_seeds):
+        add_pred = False
+        init_seed = sorted_patches[i]
+        #print('init_seed', init_seed)
         
         potentials = sorted_patches[i:k_patches+i]
-        #potentials = torch.cat((sorted_patches[:i], sorted_patches[i:k_patches+1]),0)
-        #potentials = sorted_patches
+        #print('potentials:',potentials)
+        
         #  TODO: potentials should take into consideration distance from init seed
-        similars = potentials[A[seed, potentials] > 0.0]
+        #print("sim=A[seed, potentials]:", A[init_seed, potentials])
+        similars = potentials[A[init_seed, potentials] > 0.0]
+        #print("similars:",similars)
+        #print("A[similars, :]",A[similars, :])
+
         M = torch.sum(A[similars, :], dim=0)
 
         # Box extraction
         pred, _ = detect_box(
-            M, seed, dims, scales=scales, initial_im_size=init_image_size[1:]
+            M, init_seed, dims, scales=scales, initial_im_size=init_image_size[1:]
         )
         pred = np.asarray(pred)
 
-        add_pred = aspect_ratio(pred) > 1.0 # TODO Remove assumption
+        #add_pred = aspect_ratio(pred) > 1.0 # TODO Remove assumption
         ious = 0
         
         ## TODO, if pick good iou
         ## if one bbox is completely inside another than pick the smaller one
-        if len(preds) > 0 and add_pred:
+        if len(preds) > 0:
             idx_to_remove = -1
             ious, inter, union = _bbox_iou(torch.from_numpy(pred), torch.from_numpy(np.asarray(preds)))
 
@@ -109,12 +126,16 @@ def lost(feats, dims, scales, init_image_size, k_patches=100, num_init_seeds=1,
                 ious = ious[ious!=ious[idx_to_remove]] 
 
             add_pred = add_pred and not any(ious >= iou_threshold)
+        else:
+            add_pred = True
             
         if add_pred:
             #print(ious, pred)
-            filtered_seeds.append(seed)
+            filtered_seeds.append(init_seed)
             preds.append(pred)
 
+        #raise Exception("ENDIT")  
+
     #print("Generated", len(preds), "predictions")
     return np.asarray(preds), A, scores, filtered_seeds
 

scripts/run-single-image.sh

 declare -a images=(
-    "COCO_train2014_000000000081" # SOME IMAGE NAME
+    "COCO_train2014_000000000077" # SOME IMAGE NAME
 )
 
 DATASET_PATH=$COCO_ROOT/images/train2014
@@ -9,8 +9,9 @@ OUTPUT_PATH=$LOST_PATH/outputs/samples
 
 DINO_ARCH=vit_base
 LOST_FEATURES=k
-K_PATCHES=100
+K_PATCHES=10
 PATCH_SIZE=16
+NUM_INIT_SEEDS=100
 
 rm -rf $OUTPUT_PATH
 
@@ -27,7 +28,7 @@ do
         --which_feature $LOST_FEATURES \
         --k_patches $K_PATCHES \
         --visualize pred \
-        --num_init_seeds 1
+        --num_init_seeds $NUM_INIT_SEEDS
 
     python main_lost.py \
         --image_path $DATASET_PATH/$i.jpg \
@@ -36,7 +37,7 @@ do
         --which_feature $LOST_FEATURES \
         --k_patches $K_PATCHES \
         --visualize fms \
-        --num_init_seeds 1
+        --num_init_seeds $NUM_INIT_SEEDS
 
     python main_lost.py \
         --image_path $DATASET_PATH/$i.jpg \
@@ -45,6 +46,6 @@ do
         --which_feature $LOST_FEATURES \
         --k_patches $K_PATCHES \
         --visualize seed_expansion \
-        --num_init_seeds 1
+        --num_init_seeds $NUM_INIT_SEEDS
 
 done
\ No newline at end of file

visualizations.py

@@ -140,6 +140,6 @@ def visualize_seed_expansion(image, pred, seed, pred_seed, scales, dims, vis_fol
     image[start_1:end_1, start_2:end_2, 2] = 41
 
     pltname = f"{vis_folder}/LOST_seed_expansion_{im_name}.png"
-    os.system('mkdir -p' + os.path.dirname(pltname))
+    os.system('mkdir -p ' + os.path.dirname(pltname))
     Image.fromarray(image).save(pltname)
     print(f"Image saved at {pltname}.")