update webcam&readme

README.md

@@ -21,13 +21,13 @@ To match poses that correspond to the same person across frames, we also provide
   git clone -b pytorch https://github.com/MVIG-SJTU/AlphaPose.git
   ```
 
-2. Install [pytorch](https://github.com/pytorch/pytorch)
+2. Install [pytorch 0.4.0](https://github.com/pytorch/pytorch)
   ```Shell
   chmod +x install.sh
   ./install.sh
   ```
 
-3. Download the models manually: **duc_se.pth** ([Google Drive]( https://drive.google.com/open?id=1OPORTWB2cwd5YTVBX-NE8fsauZJWsrtW) | [Baidu pan]())(2018/08/30), **yolov3.weights**([Google Drive](https://drive.google.com/open?id=1yjrziA2RzFqWAQG4Qq7XN0vumsMxwSjS) | [Baidu pan](https://pan.baidu.com/s/108SjV-uIJpxnqDMT19v-Aw)). Place them into `./models/sppe` and `./models/yolo` respectively.
+3. Download the models manually: **duc_se.pth** (2018/08/30) ([Google Drive]( https://drive.google.com/open?id=1OPORTWB2cwd5YTVBX-NE8fsauZJWsrtW) | [Baidu pan]()), **yolov3.weights**([Google Drive](https://drive.google.com/open?id=1yjrziA2RzFqWAQG4Qq7XN0vumsMxwSjS) | [Baidu pan](https://pan.baidu.com/s/108SjV-uIJpxnqDMT19v-Aw)). Place them into `./models/sppe` and `./models/yolo` respectively.
 
 
 ## Quick Start
@@ -51,6 +51,14 @@ python3 demo.py --list examples/list-coco-demo.txt --indir ${img_directory} --ou
 ```
 python3 video_demo.py --video ${path to video} --outdir examples/results/  --conf 0.5 --nms 0.45
 ```
+If you have a gpu with GPU memory larger than 8GB, consider increasing the detection batch:
+```
+python3 demo.py --indir ${img_directory} --outdir examples/res --detbatch 8
+```
+- **Note**:  If you meet OOM(out of memory) problem, decreasing the pose estimation batch until the program can run on your computer:
+```
+python3 demo.py --indir ${img_directory} --outdir examples/res --posebatch 30
+```
 - **For more**:  Checkout the [run.md](doc/run.md) for more options
 
 ## FAQ

doc/run.md

@@ -15,6 +15,8 @@ Here, we first list the flags of this script and then give some examples.
 - `--format`: The format of the saved results. By default, it will save the output in COCO-like format. An alternative option is 'cmu', which saves the results in the format of CMU-Pose. For more details, see [output.md](output.md)
 - `--conf`: Confidence threshold for human detection. Lower the value can improve the final accuracy but decrease the speed. Default is 0.2.
 - `--nms`: Confidence threshold for human detection. Increase the value can improve the final accuracy but decrease the speed. Default is 0.6.
+- `--detbatch`: Batch size for the detection network. 
+- `--posebatch`: Maximum batch size for the pose estimation network. If you met OOM problem, decrease this value until it fit in the memory.
 
 ## Examples
 - **Run AlphaPose for all images in a folder ,save the results in the format of CMU-Pose and save the images**:

install.sh

@@ -1,4 +1,4 @@
-pip3 install --user torch
+pip3 install --user torch==0.4.0
 pip3 install --user torchvision
 pip3 install --user -e git+https://github.com/ncullen93/torchsample.git#egg=torchsample
 pip3 install --user visdom

opt.py

@@ -110,7 +110,7 @@
 parser.add_argument('--mode', dest='mode',
                     help='detection mode, fast/normal/accurate', default="normal")
 parser.add_argument('--outdir', dest='outputpath',
-                    help='output-directory', default="")
+                    help='output-directory', default="examples/res/")
 parser.add_argument('--inp_dim', dest='inp_dim', type=str, default='608',
                     help='inpdim')
 parser.add_argument('--conf', dest='confidence', type=float, default=0.2,
@@ -123,16 +123,16 @@
                     help='visualize image')
 parser.add_argument('--format', type=str,
                     help='save in the format of cmu or coco or openpose, option: coco/cmu/open')
-parser.add_argument('--detbatch', type=int, default=6,
+parser.add_argument('--detbatch', type=int, default=1,
                     help='detection batch size')
 parser.add_argument('--posebatch', type=int, default=80,
                     help='pose estimation maximum batch size')
 
 "----------------------------- Video options -----------------------------"
 parser.add_argument('--video', dest='video',
                     help='video-name', default="")
-parser.add_argument('--webcam', dest='webcam',
-                    help='webcam number', default=0)
+parser.add_argument('--webcam', dest='webcam', type=str,
+                    help='webcam number', default='0')
 parser.add_argument('--save_video', dest='save_video',
                     help='whether to save rendered video', default=False, action='store_true')
 parser.add_argument('--vis_fast', dest='vis_fast',

pPose_nms.py

@@ -13,7 +13,7 @@
 mu = 1.7
 delta2 = 2.65
 gamma = 22.48
-scoreThreds = 0.1
+scoreThreds = 0.3
 matchThreds = 5
 areaThres = 0#40 * 40.5
 alpha = 0.1

webcam_demo.py

@@ -42,7 +42,7 @@ def loop():
     fvs = WebcamLoader(webcam).start()
     (fourcc,fps,frameSize) = fvs.videoinfo()
     # Data writer
-    save_path = os.path.join(args.outputpath, 'AlphaPose_'+webcam.split('/')[-1].split('.')[0]+'.avi')
+    save_path = os.path.join(args.outputpath, 'AlphaPose_webcam'+webcam+'.avi')
     writer = DataWriter(args.save_video, save_path, cv2.VideoWriter_fourcc(*'XVID'), fps, frameSize).start()
 
     # Load YOLO model
@@ -112,14 +112,17 @@ def loop():
                 ckpt_time, detNMS_time = getTime(ckpt_time)
                 runtime_profile['dn'].append(detNMS_time)
                 # Pose Estimation
-                inps, pt1, pt2 = crop_from_dets(inp, boxes)
+                inps = torch.zeros(boxes.size(0), 3, opt.inputResH, opt.inputResW)
+                pt1 = torch.zeros(boxes.size(0), 2)
+                pt2 = torch.zeros(boxes.size(0), 2)
+                inps, pt1, pt2 = crop_from_dets(inp, boxes, inps, pt1, pt2)
                 inps = Variable(inps.cuda())
 
                 hm = pose_model(inps)
                 ckpt_time, pose_time = getTime(ckpt_time)
                 runtime_profile['pt'].append(pose_time)
 
-                writer.save(boxes, scores, hm, pt1, pt2, orig_img, im_name=str(i)+'.jpg')
+                writer.save(boxes, scores, hm.cpu(), pt1, pt2, orig_img, im_name=str(i)+'.jpg')
 
                 ckpt_time, post_time = getTime(ckpt_time)
                 runtime_profile['pn'].append(post_time)