combine txt and deal with overlap pixels

Author: Spritea

RLE_check.py

@@ -0,0 +1,8 @@
+from pycocotools import mask as cocomask
+import cv2
+
+# the RLE input format for coco api.
+a={'size':[375,1242],'counts':"aaU19];3M5J6L2M1O2N1O101O00000O2O0O1O1O010O001O1O100O1O1O100000O0100O1O100O10000000001O0000001N2O001O0O100O2O000000O10O1000O100O1O1O100O1O1O1000000000000000000000000000000O1000000O100O1001O001O001O000010O01O000100O001N2N3M3N2M4J6K4Laag;"}
+b=cocomask.decode(a)
+cv2.imshow('tt',b*255)
+cv2.waitKey(0)

combine_txt_no_overlap.py

@@ -0,0 +1,91 @@
+import os
+from tqdm import tqdm
+from pycocotools import mask as cocomask
+import cv2
+import numpy as np
+
+# this file combine_txt.py is to
+# combine mots_seg_track txt outputs of each class to one file,
+# and make the overlapping pixels only belong to one object.
+# Overlap rule: assign pixels to the object with smaller area.
+
+val_in_trainval_seqmap=[2,6,7,8,10,13,14,16,18]
+seqmap=val_in_trainval_seqmap
+
+def get_RLE(line_one):
+    split_all=line_one.split(' ')
+    RLE={'size':[int(split_all[3]),int(split_all[4])],'counts':split_all[5]}
+    return RLE
+
+file_dir="Adelaidet_result/training_dir/COCO_pretrain/BoxInst_MS_R_50_1x_kitti_mots/to_mots_txt/mots_seg_track/val_in_trainval/"
+for seq_one in tqdm(seqmap):
+    car_file=file_dir+'car/'+str(seq_one).zfill(4)+'.txt'
+    pedestrian_file=file_dir+'pedestrian/'+str(seq_one).zfill(4)+'.txt'
+    file_inputs=[car_file,pedestrian_file]
+    out_dir=file_dir+'both_car_pedestrian_no_overlap/'
+    # out_dir=file_dir+'tt/'
+    out_path=out_dir+str(seq_one).zfill(4)+'.txt'
+    os.makedirs(out_dir,exist_ok=True)
+    
+    with open(car_file,'r') as f:
+        # read to a list and remove ending line character.
+        car_content=f.read().splitlines()
+    with open(pedestrian_file,'r') as f:
+        ped_content=f.read().splitlines()
+
+    # need to get the max frame id.
+    combine_content=car_content+ped_content
+    frame_list=[int(x.split(' ')[0]) for x in combine_content]
+    max_frame=max(frame_list)
+    
+    # below deal with the overlapping in one same frame.
+    combine_content_sort=sorted(combine_content,key=lambda x: int(x.split(' ')[0]))
+    new_content=[]
+    frame_overlap=[]
+    for frame_id in range(0,max_frame+1):
+        frame_one_content=list(filter(lambda x: int(x.split(' ')[0])==frame_id,combine_content_sort))
+        # skip if the frame has no mask output.
+        if not len(frame_one_content):
+            continue
+        # compute area and sort by it.
+        area_list=[]
+        for item in frame_one_content:
+            RLE=get_RLE(item)
+            area_list.append(cocomask.area(RLE))
+        objects_sort=[x for _,x in sorted(zip(area_list,frame_one_content))]
+        
+        objects_disjoint = [objects_sort[0]]
+        used_pixels = get_RLE(objects_sort[0])
+        for obj in objects_sort[1:]:
+            new_mask = get_RLE(obj)
+            # for cocomask.merge func, the argument intersect=1 gets intersection part, else gets union part,
+            # this is defined in coco official repo-common/maskApi.c.
+            if cocomask.area(cocomask.merge([used_pixels, get_RLE(obj)], intersect=True)) > 0.0:
+                obj_mask_decoded = cocomask.decode(get_RLE(obj))
+                used_pixels_decoded = cocomask.decode(used_pixels)
+                # below remove the overlapping pixels.
+                obj_mask_decoded[np.where(used_pixels_decoded > 0)] = 0
+                # have to decode bytes to string, since cocomask.encode(obj_mask_decoded) outputs bytes,
+                # and bytes starts with b', the b' at the begining has to be removed,
+                # and then its remaining part is the content could be converted to string.
+                new_mask = cocomask.encode(obj_mask_decoded)
+                new_mask['counts']=new_mask['counts'].decode('UTF-8')
+                frame_overlap.append(frame_id)
+            used_pixels = cocomask.merge([used_pixels, get_RLE(obj)], intersect=False)
+            # str(new_mask['counts']) is to change bytes type counts,
+            # generated by new_mask = cocomask.encode(obj_mask_decoded), to string.
+            obj_refine=' '.join(obj.split(' ')[:3]+[str(new_mask['size'][0]),str(new_mask['size'][1]),new_mask['counts']])
+            objects_disjoint.append(obj_refine)
+        new_content.extend(objects_disjoint)
+        
+        # print(frame_id)
+    print(f"Seq {str(seq_one).zfill(4)} has overlap frame: ", frame_overlap)
+    
+    with open(out_path,'w') as outfile:
+        for line in new_content:
+            line_split=line.split(' ')
+            if int(line_split[2])==2:
+                line_split[1]=str(int(line_split[1])+1000)
+            line_out=' '.join(line_split)+'\n'
+            outfile.write(line_out)
+