-
9/28 - 수정된 로직 결과값 분석개발일지 2024. 9. 28. 06:28
그냥 해봤다.
연산량은 많지 않은것 같은데 멀티프로세싱이 잘 안되는지 시간이 기존 방식보다 많이 걸린다.
Front와 Back만 적용했다.
약간 더 좋은 수치를 보여준다.
내가 수정한 코드 결과
mAP: 0.4622 mATE: 0.5431 mASE: 0.4506 mAOE: 0.5717 mAVE: 0.3964 mAAE: 0.2871 NDS: 0.5062 Eval time: 5.2s Per-class results: Object Class AP ATE ASE AOE AVE AAE car 0.730 0.317 0.149 0.068 0.095 0.083 truck 0.656 0.175 0.130 0.037 0.033 0.000 bus 0.741 0.521 0.124 0.071 0.457 0.052 trailer 0.000 1.000 1.000 1.000 1.000 1.000 construction_vehicle 0.000 1.000 1.000 1.000 1.000 1.000 pedestrian 0.639 0.452 0.257 0.336 0.216 0.162 motorcycle 0.684 0.404 0.292 1.084 0.055 0.000 bicycle 0.377 0.435 0.206 0.548 0.315 0.000 traffic_cone 0.797 0.127 0.349 nan nan nan barrier 0.000 1.000 1.000 1.000 nan nan기존 방식
mAP: 0.4603 mATE: 0.5421 mASE: 0.4522 mAOE: 0.5639 mAVE: 0.3982 mAAE: 0.2864 NDS: 0.5059 Eval time: 5.0s Per-class results: Object Class AP ATE ASE AOE AVE AAE car 0.740 0.312 0.150 0.067 0.095 0.078 truck 0.650 0.174 0.142 0.040 0.045 0.000 bus 0.756 0.521 0.124 0.071 0.454 0.052 trailer 0.000 1.000 1.000 1.000 1.000 1.000 construction_vehicle 0.000 1.000 1.000 1.000 1.000 1.000 pedestrian 0.639 0.454 0.255 0.333 0.217 0.161 motorcycle 0.674 0.394 0.293 1.096 0.054 0.000 bicycle 0.342 0.438 0.209 0.469 0.321 0.000 traffic_cone 0.802 0.128 0.349 nan nan nan barrier 0.000 1.000 1.000 1.000 nan nan- AP (Mean Average Precision) :
- ATE (Mean Absolute Trajectory Error) :
- ASE (Mean Absolute Scale Error) :
- AOE (Average Orientation Error)
- AVE (Average Velocity Error) :
- AAE (Average Attribute Error) : 1에서 attribute classification accuracy를 빼준 값 (1 - acc)
- NDS (nuScenes Detection Score)
- mAP, mATE, mASE, mAOE, mAVE, mAAE에 가중치를 부여하여 합산한 값
- TP 오류를 TP 점수로 변환한다
- mAP에는 가중치 5를 할당하고 나머지에는 가중치 1을 할당하여 계산한다
위 수치값은 아래 잘 설명 되어 있다
[딥러닝 프로젝트] 4. Center Point 모델 사용하여 object detection 수행하기 - 결과
centerpoint 모델 학습 결과
velog.io
정면타겟은 엄청 중요한데 수정 로직은 잘 탐지한다.
로직 수정 후
기존 로직
python3.8>site-packages>mmdet>apis>test.py>single_gpu_test
더보기def single_gpu_test(model, model_img, data_loader, show=False, out_dir=None, show_score_thr=0.3): model.eval() results = [] dataset = data_loader.dataset PALETTE = getattr(dataset, 'PALETTE', None) prog_bar = mmcv.ProgressBar(len(dataset)) model_img.eval() # BGR to RGB conversion ID_COLOR_MAP = [ (59, 59, 238), # vibrant blue (0, 255, 0), # green (0, 0, 255), # blue (255, 255, 0), # Yellow (0, 255, 255), # Cyan (255, 0, 255), # Magenta (255, 255, 255),# White (0, 127, 255), # medium sky blue (71, 130, 255), # medium cornflower blue (127, 127, 0), # Olive or dark yellow-green ] ID_COLOR_MAP_RGB = [(r, g, b) for (b, g, r) in ID_COLOR_MAP] transform_img = transforms.Compose([transforms.ToTensor()]) camera_sensors = ['CAM_FRONT', 'CAM_FRONT_RIGHT','CAM_FRONT_LEFT', 'CAM_BACK', 'CAM_BACK_LEFT','CAM_BACK_RIGHT'] target_classes = [0,1,2,3,4,6,7,8,9,10,11,12,13,14,15,16,17] for i, data in enumerate(data_loader): with torch.no_grad(): result = model(return_loss=False, rescale=True, **data) if False: raw_imgs = data['img'].data[0][0].permute(0, 2, 3, 1).cpu().numpy() # raw_imgs = data2["img"][0].permute(0, 2, 3, 1).cpu().numpy() # # 6개 카메라 이미지 합치기 imgs = [] mean = [123.675, 116.28 , 103.53 ] # np.array(cfg.img_norm_cfg["mean"]) std = [58.395, 57.12 , 57.375] # np.array(cfg.img_norm_cfg["std"]) img_norm_mean = mean#np.array(cfg.img_norm_cfg["mean"]) img_norm_std = std#np.array(cfg.img_norm_cfg["std"]) image = raw_imgs * img_norm_std + img_norm_mean for k in range(6): img = image[k] if k > 2: img = cv2.flip(img, 1) resized_img = cv2.resize(img,(1600, 900)) ubyte_img = resized_img.astype(np.uint8) # rgb_img = cv2.cvtColor(ubyte_img, cv2.COLOR_BGR2RGB) imgs.append(ubyte_img) images = np.concatenate( [ np.concatenate([imgs[2], imgs[0], imgs[1]], axis=1), np.concatenate([imgs[4], imgs[3], imgs[5]], axis=1), ], axis=0, ) # 여기까지가 출력 결과 임. # result = result[i]["img_bbox"] vis_score_threshold = 0.3 indices = [i for i, score in enumerate(result[0]["img_bbox"]["scores_3d"]) if score > 0.15] pred_bboxes_3d = result[0]["img_bbox"]["boxes_3d"][indices]#[result["scores_3d"] > vis_score_threshold] # 추가 DNN 실행하기 img_pil = Image.fromarray(images) image_tensor = transform_img(img_pil).unsqueeze(0) with torch.no_grad(): predictions = model_img(image_tensor) corners_3d = box3d_to_corners(pred_bboxes_3d) ratio_x = 1600/704 ratio_y = 900/256 num_bbox = corners_3d.shape[0] pts_4d = np.concatenate( [corners_3d.reshape(-1, 3), np.ones((num_bbox * 8, 1))], axis=-1 ) imgfov_pts_2d_CAM = [] for k, key in enumerate(camera_sensors): lidar2img_rt = copy.deepcopy(data["projection_mat"][0][k]).reshape(4, 4) if isinstance(lidar2img_rt, torch.Tensor): lidar2img_rt = lidar2img_rt.cpu().numpy() pts_2d = pts_4d @ lidar2img_rt.T pts_2d[:, 2] = np.clip(pts_2d[:, 2], a_min=1e-5, a_max=1e5) pts_2d[:, 0] /= pts_2d[:, 2] pts_2d[:, 1] /= pts_2d[:, 2] imgfov_pts_2d = pts_2d[..., :2].reshape(num_bbox, 8, 2) imgfov_pts_2d_CAM.append(imgfov_pts_2d) # # 기존에 있는 박스에 결과값 매칭하기 fastrnn_boxes = [] show_score_thr=0.5 for idx, (box, label, score) in enumerate(zip(predictions[0]['boxes'], predictions[0]['labels'], predictions[0]['scores'])): if score > show_score_thr and label in target_classes: x1, y1, x2, y2 = box box1 = [ (x1, y1), # 좌상단 (x2, y1), # 우상단 (x1, y2), # 좌하단 (x2, y2) # 우하단 ] fastrnn_boxes.append([box1,score,label,False]) # 결과값이용해서 박스 그리기 h, w = 256,704 sparse4d_boxes = [] temp_camera_sensors = ['CAM_FRONT','CAM_BACK'] for k, key in enumerate(temp_camera_sensors): if 'CAM_FRONT' == key: offset_x = 1600 offset_y = 0 elif 'CAM_FRONT_RIGHT' == key: offset_x = 3200 offset_y = 0 elif 'CAM_FRONT_LEFT' == key: offset_x = 0 offset_y = 0 elif 'CAM_BACK' == key: offset_x = 3200 offset_y = 900 elif 'CAM_BACK_LEFT' == key: offset_x = 1600 offset_y = 900 elif 'CAM_BACK_RIGHT' == key: offset_x = 4800 offset_y = 900 if 'BACK' in key: sign = -1 else: sign = 1 for ii in range(num_bbox): corners = np.clip(imgfov_pts_2d_CAM[k][ii], -1e4, 1e5).astype(np.int32) if sign > 0: box2 = [ (int(corners[2][0]*ratio_x+offset_x),int(corners[2][1]*ratio_y+offset_y)), # 좌상단 (int(corners[5][0]*ratio_x+offset_x),int(corners[5][1]*ratio_y+offset_y)), # 우상단 (int(corners[3][0]*ratio_x+offset_x),int(corners[3][1]*ratio_y+offset_y)), # 좌하단 (int(corners[4][0]*ratio_x+offset_x),int(corners[4][1]*ratio_y+offset_y)), # 우하단 ] else: box2 = [ (offset_x - int(corners[2][0]*ratio_x),int(corners[2][1]*ratio_y+offset_y)), # 좌상단 (offset_x - int(corners[5][0]*ratio_x),int(corners[5][1]*ratio_y+offset_y)), # 우상단 (offset_x - int(corners[3][0]*ratio_x),int(corners[3][1]*ratio_y+offset_y)), # 좌하단 (offset_x - int(corners[4][0]*ratio_x),int(corners[4][1]*ratio_y+offset_y)), # 우하단 ] sparse4d_boxes.append([box2,score,ii,False]) for idx1, fastrnn_box in enumerate(fastrnn_boxes): box1,score,label,checked = fastrnn_box if checked == False: for idx2, sparse4d_box in enumerate(sparse4d_boxes): box2,score2,label2,checked2 = sparse4d_box if checked2 == False: # box1 = sparse4d_boxes[8][0] # box2 = fastrnn_boxes[4][0] # 각 사각형의 좌상단 및 우하단 좌표 추출 (x1_min, y1_min), (x1_max, y1_max) = box1[0], box1[3] (x2_min, y2_min), (x2_max, y2_max) = box2[0], box2[3] # 교차 영역의 좌상단과 우하단 좌표 계산 inter_x_min = max(x1_min, x2_min) # 교차 영역의 좌상단 X 좌표 inter_y_min = max(y1_min, y2_min) # 교차 영역의 좌상단 Y 좌표 inter_x_max = min(x1_max, x2_max) # 교차 영역의 우하단 X 좌표 inter_y_max = min(y1_max, y2_max) # 교차 영역의 우하단 Y 좌표 # 교차 영역의 면적 계산 inter_area = max(0, inter_x_max - inter_x_min) * max(0, inter_y_max - inter_y_min) # 각 사각형의 면적 계산 area_box1 = (x1_max - x1_min) * (y1_max - y1_min) area_box2 = (x2_max - x2_min) * (y2_max - y2_min) # 합집합 영역의 면적 계산 union_area = area_box1 + area_box2 - inter_area # IoU 계산 if union_area > 0: iou = inter_area / union_area else: iou = 0 # 겹침 비율이 특정 임계값 이상일 때 threshold = 0.3 # 50% 겹침을 기준으로 if iou > threshold: fastrnn_boxes[idx1][3] = True sparse4d_boxes[idx2][3] = True iii = sparse4d_boxes[idx2][2] result[0]["img_bbox"]["scores_3d"][indices[iii]] = torch.tensor(0.5) break #pred_bboxes_3d = torch.cat((pred_bboxes_3d, pred_bboxes_outofspec_3d[buf_index].unsqueeze(0)), dim=0) #pred_bboxes_3d = result["boxes_3d"]#[result["scores_3d"] > vis_score_threshold] batch_size = len(result) if show or out_dir: if batch_size == 1 and isinstance(data['img'][0], torch.Tensor): img_tensor = data['img'][0] else: img_tensor = data['img'][0].data[0] img_metas = data['img_metas'][0].data[0] imgs = tensor2imgs(img_tensor, **img_metas[0]['img_norm_cfg']) assert len(imgs) == len(img_metas) for i, (img, img_meta) in enumerate(zip(imgs, img_metas)): h, w, _ = img_meta['img_shape'] img_show = img[:h, :w, :] ori_h, ori_w = img_meta['ori_shape'][:-1] img_show = mmcv.imresize(img_show, (ori_w, ori_h)) if out_dir: out_file = osp.join(out_dir, img_meta['ori_filename']) else: out_file = None model.module.show_result( img_show, result[i], bbox_color=PALETTE, text_color=PALETTE, mask_color=PALETTE, show=show, out_file=out_file, score_thr=show_score_thr) # encode mask results if isinstance(result[0], tuple): result = [(bbox_results, encode_mask_results(mask_results)) for bbox_results, mask_results in result] # This logic is only used in panoptic segmentation test. elif isinstance(result[0], dict) and 'ins_results' in result[0]: for j in range(len(result)): bbox_results, mask_results = result[j]['ins_results'] result[j]['ins_results'] = (bbox_results, encode_mask_results(mask_results)) results.extend(result) for _ in range(batch_size): prog_bar.update() return results반응형'개발일지' 카테고리의 다른 글
10/1 - 최하위 폴더 검색하기..... (2) 2024.10.01 9/22 - Sparse4D, fasterrcnn resnet50 fpn 로직 수정(feat 차량 정보) (0) 2024.09.22 9/17 - mini 데이터를 이용한 시나리오 결과 정리 (0) 2024.09.17 9/14 - NuScenes GT 정보 그리기 (7) 2024.09.14 9/8 - 차량 정보 정리 (0) 2024.09.09