search

CaptainEven / MCMOT

Real time one-stage multi-class & multi-object tracking based on anchor-free detection and ReID
MIT License
383 stars 82 forks source link

track_ids are not consistent over the video #52

Open tiru1930 opened 3 years ago

tiru1930 commented 3 years ago

track ids are keep changing over the certain number of frames for same object. These number of frames are not consistent. what could be the reason.

abhaygargab commented 3 years ago

Even i am facing the same issue. The detection results are good but the track ids keep changing frequently. @CaptainEven what are the parameters that can be tuned? Num of classes in the dataset = 5

['train.py', '--task', 'mot', '--exp_id', 'new_bboxes_3e-5', '--gpus', '1', '--batch_size', '16', '--num_epochs', '60', '--lr', '3e-5', '--lr_step', '20,40,50', '--data_cfg', '../src/lib/cfg/cater_new_bboxes.json', '--load_model', '/data/Abhay/FairMOT/models/dla34-ba72cf86.pth', '--input_h', '608', '--input_w', '1088', '--print_iter', '1', '--arch', 'dla_34'] ==> Opt: K: 200 arch: dla_34 batch_size: 16 cat_spec_wh: False chunk_sizes: [16] conf_thres: 0.4 data_cfg: ../src/lib/cfg/cater_new_bboxes.json data_dir: /mnt/diskb/even/dataset dataset: jde debug_dir: /data/Abhay/FairMOT/exp/mot/new_bboxes_3e-5/debug dense_wh: False det_thres: 0.3 down_ratio: 4 exp_dir: /data/Abhay/FairMOT/exp/mot exp_id: new_bboxes_3e-5 fix_res: True gen_scale: True gpus: [1] gpus_str: 1 head_conv: 256 heads: {'hm': 5, 'wh': 2, 'id': 128, 'reg': 2} hide_data_time: False hm_weight: 1 id_loss: ce id_weight: 1 input_h: 608 input_img: /users/duanyou/c5/all_pretrain/test.txt input_mode: video input_res: 1088 input_video: ../videos/uav_339.mp4 input_w: 1088 input_wh: (1088, 608) is_debug: False keep_res: False load_model: /data/Abhay/FairMOT/models/dla34-ba72cf86.pth lr: 3e-05 lr_step: [20, 40, 50] master_batch_size: 16 mean: None metric: loss min_box_area: 100 mse_loss: False multi_scale: True nID_dict: defaultdict(<class 'int'>, {4: 6, 1: 1, 2: 4, 3: 4, 0: 3}) nms_thres: 0.4 norm_wh: False not_cuda_benchmark: False not_prefetch_test: False not_reg_offset: False num_classes: 5 num_epochs: 60 num_iters: -1 num_stacks: 1 num_workers: 4 off_weight: 1 output_format: video output_h: 152 output_res: 272 output_root: ../results output_w: 272 pad: 31 print_iter: 1 reg_loss: l1 reg_offset: True reid_cls_ids: 0,1,2,3,4 reid_dim: 128 resume: False root_dir: /data/Abhay/FairMOT save_all: False save_dir: /data/Abhay/FairMOT/exp/mot/new_bboxes_3e-5 seed: 317 std: None task: mot test: False test_mot15: False test_mot16: False test_mot17: False test_mot20: False track_buffer: 30 trainval: False val_intervals: 10 val_mot15: False val_mot16: False val_mot17: False val_mot20: False vis_thresh: 0.5 wh_weight: 0.1

abhaygargab commented 3 years ago

As the objects in my videos start and stop moving suddenly, that too at a high speed. So can the Kalman filtering be a reason for frequent ID switches?

If yes, then is there a way to stop using Kalman filters during inferencing?

ThankYou