Output all bbox

VidSTG/read_annotation.py

@@ -2,6 +2,7 @@ import json
 import os
 import cv2
 from tqdm import tqdm
+import re
 
 from generate_features import write_npy
 
@@ -23,10 +24,10 @@ def normalize_bbox(bbox, width, height):
     Normalize the bbox
     """
     xmin, ymin, xmax, ymax = bbox
-    xmin = round(xmin / width, 3)
-    ymin = round(ymin / height, 3)
-    xmax = round(xmax / width, 3)
-    ymax = round(ymax / height, 3)
+    xmin = int(round(xmin / width, 2) * 100)
+    ymin = int(round(ymin / height, 2) * 100)
+    xmax = int(round(xmax / width, 2) * 100)
+    ymax = int(round(ymax / height, 2) * 100)
     
     return [xmin, ymin, xmax, ymax]
 
@@ -81,16 +82,19 @@ def generate_prompt(vid, vid_path, question, answer,
 
     target = vido_data['subject/objects'][question_tid]['category']
 
-    start_bbox = vido_data['trajectories'][temporal_gt_begin_fid]
-    found = False
-    for a_dict in start_bbox:
-        if question_tid is a_dict['tid']:
-            start_bbox = a_dict['bbox']
-            found = True
-            break
-    if not found:
-        raise ValueError("start_bbox not found")
-    start_bbox = list(start_bbox.values())
+    frame_to_bbox = {}
+    for fid in range(temporal_gt_begin_fid, temporal_gt_end_fid + 1):
+        start_bbox = vido_data['trajectories'][fid]
+        found = False
+        for a_dict in start_bbox:
+            if question_tid is a_dict['tid']:
+                start_bbox = a_dict['bbox']
+                found = True
+                break
+        if not found:
+            raise ValueError("start_bbox not found")
+        start_bbox = list(start_bbox.values())
+        frame_to_bbox[fid] = start_bbox
     
     # # verify the start_bbox
     # cap = cv2.VideoCapture(vid_path)
@@ -105,26 +109,6 @@ def generate_prompt(vid, vid_path, question, answer,
     # cv2.imshow('Frame', frame)
     # cv2.waitKey(0)
     
-    found = False
-    end_bbox = vido_data['trajectories'][temporal_gt_end_fid]
-    for a_dict in end_bbox:
-        if question_tid is a_dict['tid']:
-            end_bbox = a_dict['bbox']
-            found = True
-            break
-    if not found:
-        raise ValueError("end_bbox not found")
-    end_bbox = list(end_bbox.values())
-    
-    # # verify the end_bbox
-    # cap.set(cv2.CAP_PROP_POS_FRAMES, int(temporal_gt_end_fid))
-    # ret, frame2 = cap.read()
-    # cv2.rectangle(frame2, (end_bbox[0], end_bbox[1]), 
-    #                         (end_bbox[2], end_bbox[3]), (0, 0, 255), 1)
-    # cv2.imshow('Frame', frame2)
-    # cv2.waitKey(0)
-    # cv2.destroyAllWindows()
-    
     
     width = vido_data['width']
     height = vido_data['height']
@@ -147,33 +131,30 @@ def generate_prompt(vid, vid_path, question, answer,
         frame_count = original_frame_count
         
     # normalized start and end frame
-    start_frame = temporal_gt_begin_fid - begin_fid
-    if asked_frames != 0:
-        start_frame = asked_frames * (start_frame / original_frame_count)
-        start_frame = int(start_frame)
-        
-    end_frame = temporal_gt_end_fid - begin_fid
-    if asked_frames != 0:
-        end_frame = asked_frames * (end_frame / original_frame_count)
-        end_frame = int(end_frame)
-        
-    write_npy(vid_path, vid, begin_fid, end_fid, asked_frames)
+    normalize_frame_to_bbox = {}
+    for fid in range(temporal_gt_begin_fid, temporal_gt_end_fid + 1):
 
+        relative_start_fid = fid - begin_fid
+        if asked_frames != 0:
+            normalized_frame = asked_frames * (relative_start_fid / original_frame_count)
+            normalized_frame = int(normalized_frame)
+            
+        normalize_frame_to_bbox[normalized_frame] = normalize_bbox(frame_to_bbox[fid], width, height)
     
-    # normalize the bbox
-    start_bbox = normalize_bbox(start_bbox, width, height)
-    end_bbox = normalize_bbox(end_bbox, width, height)
-    
-    new_prompt = r' The {} is initially at {} in frame {} and stops appearing at position {} in frame {}'
+    write_npy(vid_path, vid, begin_fid, end_fid, asked_frames)
+
+    new_prompt = r' The {} is at '.format(target)
+    bboxes = re.sub(r'\s+', '', str(normalize_frame_to_bbox))
+    new_prompt += bboxes
     json_obj = {}
     json_obj["id"] = vid
     
     question_dict = {"from": "human", "value": "<video>\n"+question['description']}
-    answer_dict = {"from": "gpt", "value": answer['description'] + new_prompt.format(target, "<b0>", "<s0>", "<b1>", "<e0>")}
+    answer_dict = {"from": "gpt", "value": answer['description'] + new_prompt}
     json_obj["conversations"] = [question_dict, answer_dict]
     
-    token = {"<s0>": start_frame, "<e0>" : end_frame, "<b0>": start_bbox, "<b1>": end_bbox}
-    json_obj["meta"] = {"token": token, "duration": frame_count}
+    # token = {"<s0>": start_frame, "<e0>" : end_frame}
+    json_obj["meta"] = {"duration": frame_count}
     
     return json_obj
 
@@ -319,6 +300,9 @@ def process_record(record, vidor_anno_path_base, vidor_path_base):
 
         if result is not None:
             all_results.append(result)
+        # print(json.dumps(result, indent=4))
+        # visualize_video(vid_path, begin_fid, end_fid, temporal_gt_begin_fid, temporal_gt_end_fid,
+        #                 question_tid, answer_tid, vido_data)
             
     return all_results
             
@@ -349,9 +333,6 @@ def main():
         if results is None:
             continue
         all_results.extend(results)
-        # print(json.dumps(result, indent=4))
-        # visualize_video(vid_path, begin_fid, end_fid, temporal_gt_begin_fid, temporal_gt_end_fid,
-        #                 question_tid, answer_tid, vido_data)
         
         if index % 100 == 0:
             with open('results.json', 'w') as json_file:

run_vtimellm.md

@@ -1,3 +1,5 @@
+
+### run example inference
 python -m vtimellm.inference --model_base lmsys/vicuna-7b-v1.5 \
 --pretrain_mm_mlp_adapter checkpoints/vicuna-7b-v1.5/vtimellm-vicuna-v1-5-7b/vtimellm-vicuna-v1-5-7b-stage1/mm_projector.bin \
 --stage2 checkpoints/vicuna-7b-v1.5/vtimellm-vicuna-v1-5-7b/vtimellm-vicuna-v1-5-7b-stage2 \
@@ -9,7 +11,7 @@ python demo_gradio.py --model_base lmsys/vicuna-7b-v1.5 \
 --stage2 ../checkpoints/vicuna-7b-v1.5/vtimellm-vicuna-v1-5-7b/vtimellm-vicuna-v1-5-7b-stage2 \
 --stage3 ../checkpoints/vicuna-7b-v1.5/vtimellm-vicuna-v1-5-7b/vtimellm-vicuna-v1-5-7b-stage3
 
-
+### port forwarding
 ssh -t -t -i /home/datasets/xitong_id_rsa xitong@newton.ist.ucf.edu -L 7860:localhost:7860 ssh evc23 -L 7860:localhost:7860
 
 ### generate validation datasets
@@ -24,6 +26,12 @@ python Shikra_V/VidSTG/read_annotation_multithread.py --vidstg VidSTG-Dataset/an
 python Shikra_V/VidSTG/read_annotation_multithread.py --vidstg VidSTG-Dataset/annotations/test_annotations.json --vidor_anno_path_base vidor/validation_annotation/validation --vidor_path_base vidor/validation/video   
 ```
 
+### Calculate the iou by using test datasets
+```
+python vtimellm/eval/eval.py --stage3 checkpoints/vtimellm-vicuna-v1-5-7b-stage3_xl_300_epoch/checkpoint-1700 --data_path data/xl/test/results_test.json --feat_folder data/xl/test/stage4_features_test --log_path vtimellm/eval/log/iou.txt --task iou
+
+```
+
 ### Verify by using my trained stage2
 
 ```
@@ -44,4 +52,4 @@ python demo_gradio.py --model_base lmsys/vicuna-7b-v1.5 \
 
 ### status
 
-we have generated 44087 training samples and 4892 validation samples
+we have generated 44087 training samples, 4892 validation samples and 5655 test samples