Update README.md

README.md

@@ -62,29 +62,26 @@ task_categories:
 ```bash
 pip install -r requirements.txt 
 ```
-##### 2. Command Line Arguments
+##### 2. Multi-GPU Support
+How Multi-GPU Works
+- Videos are put into a task queue.
+- Each GPU process take one task from the queue when vacant.
+- If failed, the task will be put back into the queue.
+- Progress bars show accumulation for all results.
+- Every time when a task is finished, the result file is updated. You can obtain intermediate results from the file.
 
+The metrics computation supports multi-GPU parallel processing for faster evaluation.
+```bash
+python src/process.py --gt_root /path/to/MIND-Data --test_root /path/to/test/videos --num_gpus 8 --metrics lcm,visual,action
+```
 - `--gt_root`: Ground truth data root directory (required)
 - `--test_root`: Test data root directory (required)
 - `--dino_path`: DINOv3 model weights directory (default: `./dinov3_vitb16`)
 - `--num_gpus`: Number of GPUs to use for parallel processing (default: 1)
 - `--video_max_time`: Maximum video frames to process (default: `None` = use all frames)
 - `--output`: Output JSON file path (default: `result_{test_root}_{timestamp}.json`)
-- `--metrics`: Comma-separated metrics to compute (default: `lcm,visual,dino,action`)
-##### 3. Multi-GPU Support
-
-The metrics computation supports multi-GPU parallel processing for faster evaluation.
-
-```bash
-python src/process.py --gt_root /path/to/MIND-Data --test_root /path/to/test/videos --num_gpus 8 --metrics lcm,visual,action
-```
-**How Multi-GPU Works**
-- Videos are put into a task queue.
-- Each GPU process take one task from the queue when vacant.
-- If failed, the task will be put back into the queue.
-- Progress bars show accumulation for all results.
-- Every time when a task is finished, the result file is updated. You can obtain intermediate results from the file.
-##### 4. How to order your test files
+- `--metrics`: Comma-separated metrics to compute (default: `lcm,visual,dino,action,gsc`)
+##### 3. How to order your test files
 ```
 {model_name}
 ├── 1st_data
@@ -94,8 +91,12 @@ python src/process.py --gt_root /path/to/MIND-Data --test_root /path/to/test/vid
 |   |   ...
 |   |    
 │   ├── mirror_test
-│   |   ├── {corresponding data name}
-│   |   │   └── video.mp4
+│   |   ├── {arbitrary data name}
+│   │   │   ├── path-1.mp4
+│   │   │   ├── path-2.mp4
+│   │   │   ├── path-3.mp4
+│   │   │   ...
+│   │   │   └── path-10.mp4
 |   |   ...
 |   |
 |   └── mem_test
@@ -110,8 +111,12 @@ python src/process.py --gt_root /path/to/MIND-Data --test_root /path/to/test/vid
 |   |   ...
 |   |    
 │   ├── mirror_test
-│   |   ├── {corresponding data name}
-│   |   │   └── video.mp4
+│   |   ├── {carbitrary data name}
+│   │   │   ├── path-1.mp4
+│   │   │   ├── path-2.mp4
+│   │   │   ├── path-3.mp4
+│   │   │   ...
+│   │   │   └── path-10.mp4
 |   |   ...
 |   |
 │   └── mem_test
@@ -123,7 +128,7 @@ python src/process.py --gt_root /path/to/MIND-Data --test_root /path/to/test/vid
 - `{corresponding data name}`: corresponding ground truth data file name
 
 
-#####  5. The detailed information of output **<span style="color:red">`Result.json`</span>**
+#####  4. The detailed information of output **<span style="color:red">`Result.json`</span>**
 
 ```
 {
@@ -137,6 +142,26 @@ python src/process.py --gt_root /path/to/MIND-Data --test_root /path/to/test/vid
       "mark_time": [int] the divider of memory context and expected perdiction; the start frame index of the expected prediction.
       "total_time": [int] the total frames of the ground truth video.
       "sample_frames": [int ]the total frames of the video to be tested.
+      "video_results": [ the general scene consistency metric result.
+        {
+          "video_name": [string] the name of the video of the specific action path
+          "error": [string] the error occur when computing metrics in this video
+          "mark_time": [int] the divider of prediction and mirror perdiction; the start frame index of the mirror prediction.
+          "sample_frames": [int] total frames of prediction and mirror perdiction; should be 2x of marktime.
+          "gsc": { 
+            "length": [int] length of the origin prediction and the mirror prediction.
+            "mse": [list[float]] the per-frame mean square error.
+            "avg_mse": [float] the average of mse.
+            "lpips": [list[float]] the per-frame Learned Perceptual Image Patch Similarity.
+            "avg_lpips": [float] the average of lpips.
+            "ssim": [list[float]] the per-frame Structural Similarity Index Measure.
+            "avg_ssim": [float] the average of ssim.
+            "psnr": [list[float]] the per-frame Peak Signal-to-Noise Ratio.
+            "avg_psnr": [float] the average of psnr.
+          }
+        },
+        ...
+      ]
       "lcm": { the long context memory metric result.
         "mse": [list[float]] the per-frame mean square error.
         "avg_mse": [float] the average of mse.
@@ -209,18 +234,20 @@ MIND-Data
 ├── 1st_data
 │   ├── test
 │   │   ├── action_space_test
-│   │   │   ├── data-{i}
+│   │   │   ├── {gt data name}
 │   │   │   │   ├── action.json
+│   │   │   │   ├── images.txt
 │   │   │   │   └── video.mp4
 |   |   |   ...
 |   |   |    
 │   │   └── mem_test
-│   │       ├── data-{i}
+│   │       ├── {gt data name}
 │   │       │   ├── action.json
+│   │       │   ├── images.txt
 │   │       │   └── video.mp4
 |   |       ...
 |   └── train
-|       ├── data-{i}
+|       ├── {gt data name}
 |       │   ├── action.json
 |       │   └── video.mp4
 |       ...
@@ -228,18 +255,20 @@ MIND-Data
 ├── 3rd_data
 │   ├── test
 │   │   ├── action_space_test
-│   │   │   ├── data-{i}
+│   │   │   ├── {gt data name}
 │   │   │   │   ├── action.json
+│   │   │   │   ├── images.txt
 │   │   │   │   └── video.mp4
 |   |   |   ...
 |   |   |    
 │   │   └── mem_test
-│   │       ├── data-{i}
+│   │       ├── {gt data name}
 │   │       │   ├── action.json
+│   │       │   ├── images.txt
 │   │       │   └── video.mp4
 |   |       ...
 |   └── train
-|       ├── data-{i}
+|       ├── {gt data name}
 |       │   ├── action.json
 |       │   └── video.mp4
 |       ...